OpenAI is reportedly preparing to launch a new subscription tier called ChatGPT Pro Lite, priced at $100 per month. Developer Tibor Blaho discovered references to the plan in the frontend code of the ChatGPT web app. OpenAI has not confirmed the plan or shared details about its release date, but the discovery suggests the company is working to address a pricing gap in its lineup.

OpenAI currently offers several subscription plans, including Free, Go ($8/month), Plus ($20/month), Pro ($200/month), Team, Business, and Enterprise options. Many users have criticized the large jump from Plus to Pro, leaving power users without a middle-ground option. The Pro Lite plan could fill this gap by targeting freelancers, researchers, and developers who require higher usage quotas but cannot justify Pro’s $200 price tag.

The new plan is expected to provide more advanced features than Plus but fewer capabilities than Pro. Code references suggest Pro Lite could offer three to five times the deep reasoning model usage quota of the Plus plan. Additionally, it may be designed to accommodate Codex usage.

This potential move comes as OpenAI navigates a challenging 2026. Despite $20 billion in revenue, the company faces a projected $14 billion annual loss due to ballooning compute costs, prompting a desperate search for a $100 billion funding lifeline.

Meanwhile, competition is heating up. The Gemini 3 now commands over 20% of the market thanks to Android integration, and Claude 4 has overtaken OpenAI in enterprise spending. User backlash over the abrupt retirement of the GPT-4o model has only added to the turbulence.

Pro Lite could be a game-changer, offering a more affordable option for power users while positioning OpenAI competitively against rivals like Google, which lacks a comparable mid-tier offering. Whether this will be enough to stabilize OpenAI remains to be seen.

In related news, OpenAI may also launch a ChatGPT-powered smart speaker equipped with a camera.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

(Source, Via)

The post OpenAI may launch new $100/month ChatGPT Pro Lite subscription appeared first on Gizmochina.

Google has introduced Gemini 3.1 Pro, the latest version of its advanced AI model. The update delivers significant improvements in reasoning and problem-solving, making it one of the most powerful AI systems available today. The model is now rolling out in preview for developers, enterprises, and consumers, with wider availability expected soon.

Gemini 3.1 Pro focuses on handling complex, multi-step tasks that require advanced reasoning. The model can build interactive experiences, translate creative ideas into functional code, and connect complex APIs with user-friendly interfaces.

It can generate immersive projects such as 3D simulations with real-time interaction, create live data dashboards like ISS orbit visualizations, and turn literary themes into modern web designs.

The model also generates animated SVGs directly from text prompts. These animations are created as code rather than video files, which preserves image quality at any scale while keeping file sizes smaller.

In terms of performance, Gemini 3.1 Pro achieved a score of 77.1% on the ARC-AGI-2 benchmark, which tests abstract reasoning. This result more than doubles the score of its predecessor, Gemini 3 Pro, and outperforms competitors like GPT-5.2 (52.9%) and Opus 4.6 (68.8%). The model also posted strong results on GPQA Diamond, SWE Bench Verified, and BrowseComp, which evaluate scientific reasoning, coding ability, and search-related tasks.

Google is making Gemini 3.1 Pro available through several platforms. Developers can access it via Google AI Studio, Gemini CLI, and Android Studio. Enterprises can use it through Vertex AI and Gemini Enterprise. Consumers can try it on the Gemini app and NotebookLM, with Pro and Ultra subscribers receiving higher usage limits.

Learn more about the model and how to access it with these top five ways to use Gemini 3.1 Pro for free.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

The post Google releases Gemini 3.1 Pro, pushing deeper into the AI reasoning race appeared first on Gizmochina.

Gemini 3.1 Pro is the company’s most capable LLM with stronger core reasoning and better performance on complex, multi-step work. It outperforms the best from Anthropic and OpenAI across many benchmarks. Here, I’ll show you the top 5 ways to access Gemini 3.1 Pro for free with all the steps.

How to use Gemini 3.1 Pro for free (5 methods)

1. Google Gemini

This is the easiest and quickest way to test the capabilities of Gemini 3.1 Pro. Simply go to gemini.google.com, choose the “Pro” option, and get things done.

Step 1: Open gemini.google.com on your device.

Step 2: Log in to your Google account (if not logged in).

Step 3: Tap the mode selector and choose “Pro”.

Step 4: Type your prompt and hit Enter.

If you’re using a smartphone, it’s best to download the Google Gemini app from Google Play Store (Android) or App Store (iPhone). The steps remain the same.

Please note that the free version of Google Gemini allows only a few messages per day for Gemini 3.1 Pro. That’s best for the testing purpose. However, you can upgrade to the Pro or Ultra plan to increase your limits. Luckily, there are ways to get Google Gemini Pro for free for a year. Comment below if you want a detailed post on that.

2. Google’s Vertex AI platform

Google is offering $300 in free credits to use Google Cloud services, including the Vertex AI platform, which allows us access to Gemini 3.1 Pro, Google Veo, and many other AI tools. Note that free credits are available for 90 days. After which, they expire.

Here’s how to sign up for Google Cloud:

Step 1: Open the Google Cloud trial offer page and tap “Get started for free”.

Step 2: Choose your Google account and your country.

Step 3: Fill in your contact information, then verify a payment method.

[Note: You won’t be charged automatically even after the free trial ends unless you do a manual activation. You can verify the same in the screenshot below.]

Step 4: In your Google Cloud dashboard, tap the menu button and choose “Vertex AI”.

Step 5: Click “Try now” below Vertex AI Studio.

Step 6: Click “Agree & continue” to enable the required APIs for using Vertex AI Studio.

Step 7: Choose “Gemini 3.1 Pro” and start using it.

The $300 free credit gives you very high usage limits to Google’s top AI models for 90 days. You can even create videos, build tools using coding, and do a lot more stuff without paying a penny.

3. Try That LLM

This is an easy and free tool for comparing responses from various LLMs. The platform offers dozens of LLMs to test your prompt, including Gemini 3.1 Pro, Claude Opus 4.6, and ChatGPT 5.2.

This is super easy to use. Just sign up on the website (it takes less than a minute), choose your model, and start using it. The best part is you don’t need to fill in payment details.

Step 1: Go to trythatllm.com and click the “Sign up” button.

Step 2: Fill in basic details to create your account.

Step 3: Verify your email.

Step 4: Tap “New Prompt Group”.

Step 5: Click “Choose an LLM” and select Google > Gemini 3.1 Preview.

Step 6: Click “Confirm LLM + Start Prompting”.

Step 7: Enter your message in the prompt section and click “Confirm Prompt”.

You get 1000 free credits when you create an account. Note that these are one-time credits and are not added every month.

4. Arena AI

Arena AI also lets you compare different LLMs for free. The best part is that you don’t need to sign up on the platform to use Gemini 3.1 Pro or any other LLM. Just open the website, choose your AI model, and start messaging.

Step 1: Go to Arena.ai and make sure side-by-side mode is selected.

Step 2: Tap on the first LLM field.

Step 3: Search for “Gemini 3.1 Pro” and choose the right LLM from.

Step 4: Write your message/prompt in the Ask anything section and enter.

Arena AI is completely free to test LLMs, but it doesn’t offer unlimited responses. The platform provides pretty generous limits, which should be more than enough for small use cases.

5. ChatPlayGround AI

ChatPlayGround AI also offers access to dozens of LLM models, including Gemini 3.1 Pro, Claude Opus 4.6, and ChatGPT 5.2. Unlike Try That LLM and Arena AI, this AI tool lets you use images and PDFs with your message. But what’s the catch here? It doesn’t have a free plan; instead, it offers a 3-day free trial, and for that, you’ll need to link your bank cards.

So, if you only need access to Gemini 3.1 Pro and other LLMs with better capabilities, you can go for the free trial, but make sure to cancel the plan before three days to avoid any charges. Simply go to chatplayground.ai, tap “Sign Up”, and fill in the required details to get started.

Be sure to visit the AI section on Gizmochina regularly for the latest updates.

The post Use Gemini 3.1 Pro for free: Here are the top 5 ways appeared first on Gizmochina.

The Internet is fuming as OpenAI retired GPT-4o recently, a chatbot model beloved for its warm, conversational style and unique emotional resonance. Despite a Change.org petition amassing over 22,000 signatures, OpenAI stood firm in its decision, citing low usage (just 0.1% of users) and the evolution of its newer models, GPT-5.1 and GPT-5.2.

This isn’t the first time OpenAI has tried to sunset GPT-4o. In 2025, the company faced overwhelming backlash from users who described the model as irreplaceable. OpenAI reversed the retirement within 24 hours, but this time, it seems the shutdown is final, leaving thousands of fans feeling abandoned.

For many, GPT-4o was more than a tool, it was a companion. Social media is flooded with grief-stricken posts likening its retirement to losing a close friend. Paying subscribers have threatened to cancel their memberships, and a few have already cancelled in protest, with one X user declaring, “No 4o, no money.”

The petition, spearheaded by Sophie Witt in April 2025, called on OpenAI to recognize GPT-4o’s unique appeal, which transcends performance metrics. Witt urged supporters to amplify their voices online, sparking widespread discussions about the emotional bond users form with AI.

However, GPT-4o wasn’t without controversy. At least eight lawsuits accuse the model of exacerbating mental health crises, with claims that its affirming tone contributed to self-harm in vulnerable users. These cases highlight the ethical challenges of emotionally intelligent AI, raising questions about where to draw the line between support and dependency.

OpenAI acknowledged these complexities in its January blog post, noting that feedback from GPT-4o users shaped its newer models. Yet, the decision to retire the model reflects a broader shift toward prioritizing safety and scalability over nostalgia.

The GPT-4o saga offers a clear lesson for marketers and AI developers about the risks and rewards of emotional design. As AI systems play a growing role in human interaction, the need to balance technological progress with accountability has never been more pressing.

In related AI news, Google could allow websites to opt out of AI Overviews following regulatory pressure in the UK.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

(Via)

The post Over 22,000 Sign Petition to Save GPT-4o as OpenAI Retires Popular Model appeared first on Gizmochina.

Google is exploring new options to allow websites and publishers to opt out of having their content used in AI Overviews. The move comes after the UK government urged the company to provide more control to content creators, citing growing concerns about fair usage and declining traffic to news sites.

UK regulators demand more control for publishers

According to a Reuters report, the UK’s Competition and Markets Authority (CMA) has opened a formal consultation on how Google manages content across its AI-powered features. The regulator wants the company to let websites opt out of having their material scraped for AI-generated summaries or used to train separate AI models like Gemini.

In a blog post published on Thursday, Google confirmed that it is working on updates that would enable websites to exclude their content from Search generative AI features. The company said it aims to protect the usefulness of Search while also giving publishers the tools to manage how their content appears across AI products.

Google introduced AI Overviews in mid-2024. The feature uses generative AI to provide quick answers at the top of search results. However, publishers have not had any mechanism to opt out of these summaries, which has led to significant drops in click-through traffic.

The CMA stated that Google holds more than 90% of the UK’s search market. It warned that this dominance could distort competition and harm news organizations that rely on visibility in search.

Google said any new opt-out control must be scalable and not create a fragmented or confusing experience for users. The company also pointed to existing tools like robots.txt and Google-Extended, but acknowledged that further updates are needed.

The CMA will spend the next year reviewing the consultation. Google said it hopes to work with publishers and regulators to find a solution that benefits both users and content creators.

In related news, you can also read our recent article exploring whether smartphones have truly reached peak innovation or if progress is now happening behind the scenes.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

The post Google may let websites opt out of AI Overviews after UK pressure appeared first on Gizmochina.

Smartphone use now exceeds three hours a day on average, while total daily screen time for many adults crosses six hours. This constant close-up focus has made eye fatigue, dryness, blurred vision, and headaches a common part of modern life. As screens become unavoidable, interest is growing in technologies that don’t just relieve symptoms but train the eyes to cope better with digital habits.

Eyeary and the Technology Behind It

South Korea–based Edenlux is preparing to launch Eyeary, a lightweight, glasses-style vision training device, in the U.S. later this year. Unlike software-only eye exercises or blue-light filters, Eyeary works directly on the ciliary muscle, which controls how the eye focuses. Heavy screen use keeps this muscle locked in a contracted state. Eyeary repeatedly guides the eyes through controlled focus changes using 144 focal points, training the muscle to relax and contract more naturally.

Is This Technology Truly New?

Eye training tools are not new, but most existing solutions are either clinical devices, simple exercise apps, or bulky headsets. Eyeary stands out by combining hardware-based visual training, AI-driven personalization, and everyday wearability. Few consumer devices currently offer this level of precision in focus training outside clinical environments.

Current Technology:

Blue-Light and Anti-Glare Glasses

Most products on the market today are blue-light filtering or anti-glare glasses. These are designed to reduce screen glare and improve viewing comfort during long hours of phone or computer use. While they are widely adopted, scientific evidence on their ability to reduce digital eye strain remains mixed.

Basic Eye Exercise Tools

Some products focus on simple eye exercises, including pinhole glasses and manual vision training tools. These aim to encourage eye movement and focus changes but lack precision. They do not use sensors, software, or personalization, and results often depend heavily on user consistency.

Clinical and Experimental Solutions

More advanced vision training exists mainly in clinical or research environments. These include VR-based vision therapy systems and adjustable-focus lenses. However, such solutions are usually bulky, expensive, or designed for supervised use rather than everyday consumers.

Market Gap in Active Vision Training

Overall, most current solutions are passive and symptom-focused. Consumer-ready, wearable devices that actively train eye-focusing muscles using data and AI are still rare, leaving this segment largely unexplored in the mainstream market.

Potential Impact on Eye Health

If effective at scale, Eyeary could shift eye care from treatment to daily prevention and recovery, especially for heavy screen users. It signals a future where vision health becomes part of consumer technology, much like fitness tracking, potentially influencing how smartphones, wearables, and displays are designed to protect long-term eye function.

Preview	Product
	LENSKART BLU | Zero Power Blue Cut Computer Glasses | Gaming Glasses | Anti Glare, Lightweight &…	Check Price on Amazon

Please Note: When you buy something using the links in our articles, we may earn a small commission at no cost to you.

Read More:

• Honor Is Going All-In on Laptops in 2026: Massive PC Lineup Leak Reveals Big Plans
• Bill Gates Backed Startup Creates Light-Powered AI Chip That Can Beat Nvidia
• NVIDIA Caught in US- China Crossfire: Jensen Huang Lands in Shanghai amidst Ban on H200 Chips

(via)

The post Screen Addiction Is Hurting Your Eyes ?: AI-Powered Vision Training Glasses Can Help You appeared first on Gizmochina.

OpenAI has launched ChatGPT Go globally, expanding its affordable subscription plan to all markets where ChatGPT is available. The plan costs $8 per month in the US and sits between the free tier and the $20 per month ChatGPT Plus subscription. ChatGPT Go was first introduced in India in August 2025 and later expanded to over 170 countries.

ChatGPT Go gives users access to GPT-5.2 Instant, OpenAI’s faster and lighter model. The plan includes significantly higher usage limits than the free tier, including more messages, file uploads, and image generation. OpenAI has not shared exact figures for these limits but says Go provides roughly ten times the allowance of the free plan. ChatGPT Go also includes a longer context window and improved memory support compared to the free version.

OpenAI confirmed that it will begin testing ads in the Free and Go tiers in the US. Ads will appear below ChatGPT’s responses and will be labeled clearly as “Sponsored.” The company said advertisers cannot influence chatbot responses and that it will not share user conversations with advertisers. Ads will not appear next to content about sensitive or regulated topics, and users under 18 will not see ads.

CEO Sam Altman said the ad rollout aims to help OpenAI keep ChatGPT affordable. He acknowledged past criticism of ads but said this approach is necessary to support a growing number of users who want frequent access to AI without paying high fees. Altman cited Instagram’s ad experience as an inspiration for the format OpenAI will use.

ChatGPT Plus and Pro users will continue to see an ad-free experience. OpenAI said it will monitor feedback and make adjustments as the ad trial progresses. The company reported strong adoption of ChatGPT Go in earlier markets and expects the global launch to accelerate usage further.

In related news, OpenAI is reportedly planning AI-powered SweetPea earbuds, positioning them as a potential alternative to Apple’s AirPods.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

(Source)

The post ChatGPT Go goes global: OpenAI bets on ads to make AI affordable appeared first on Gizmochina.

GUANGZHOU, China — Held in Guangzhou from December 11, the 2025 Global Technology Innovation Conference (TIC) served as the flagship stage for the company to unveil its latest advancements in integrated technology. The event centered on the core theme of “AI for Real,” showcasing how artificial intelligence is being woven into every layer of TCL’s ecosystem—from consumer-facing smart products and immersive user experiences to the very backbone of its sophisticated manufacturing processes.

At the conference, a clear message emerged from one of the world’s largest consumer electronics makers: artificial intelligence is not just a buzzword—it’s a tool for tangible, vertical integration. In an exclusive interview with Gizmochina, Daniel Sun, CTO of TCL Industries, shed light on how the company is weaving AI into everything from TVs and air conditioners to its own advanced display factories, all while staying grounded in what he calls “experience-first” technology.

Beyond the Hype: AI as an Enabler, Not a Magic Wand

Sun began by addressing widespread misconceptions about AI. “For consumers, intelligence might mean using an app to turn off an air conditioner remotely,” he said. “For engineers, that’s simply Wi‑Fi, apps, and cloud services.” This gap, he argues, is where many tech companies lose sight of the real goal: delivering meaningful experiences rather than chasing the latest model for its own sake.

He emphasized that not every product needs—or can run—a large language model (LLM). “Many problems can be solved with small models,” Sun explained, pointing to quality inspection in TCL CSOT’s display manufacturing. “Using an AI device with a camera to inspect panel circuits is like an airport X-ray scan. It used to be manual; now it’s fully automated—and it doesn’t require an LLM.”

Why TCL CSOT Built Its Own AI: The X-Intelligence 3.0 Story

When asked about TCL CSOT’s homegrown AI system, X-Intelligence 3.0, Sun outlined two core motivations. First, public LLMs lack deep, proprietary knowledge. “We attached our specialized display materials data to the world’s best LLMs, and the results were surprisingly poor,” he said. “The model simply wasn’t pretrained on this domain.” In other words, they couldn’t comprehend highly specialized, cutting-edge knowledge.

Instead of building from scratch, TCL CSOT leverages its vast internal knowledge—accumulated over years of R&D and billions in investment—into existing high-performance architectures through fine-tuning and reinforcement learning. “We leverage the best open models and enhance them with our private expertise,” Sun noted, citing cooperative and open‑source advances, including those inspired by frameworks like DeepSeek.

The second driver is organizational efficiency. Sun pointed out that in corporations with thousands of employees, decision-making processes tend to decelerate. He emphasized that while AI cannot instantly replicate human agility, it serves as a powerful accelerator in domains such as operational analysis, process optimization, and high-value task automation — a view supported by its growing role in enhancing enterprise workflow efficiency.

From Living Rooms to Factories: AI in Action

On the product side, TCL is taking an open‑ecosystem approach. In the U.S., TCL TVs now integrate Google’s Gemini, enabling voice-activated content discovery, educational YouTube browsing, and smart‑home control. In China, the company uses cloud-based LLMs—quantized and cost-optimized—to power natural language TV interfaces for under a dollar per device per year.

But AI isn’t just for the end user. In manufacturing, LLMs and small models work in tandem. At the TCL CSOT Guangzhou t9 factory, which Gizmochina toured on December 11, AI assists in everything from panel defect detection to production scheduling. The t9 line—the world’s only large-generation line compatible with LCD, Micro LED, and inkjet-printed OLED—relies on AI to maintain precision across its 180,000 glass substrates monthly capacity. Capable of producing 6- to 100-inch panels, the t9 line enables applications spanning smartphones, tablets, notebooks, vehicle displays, monitors, TVs, and large-format commercial screens.

The Hidden AI Star: Energy‑Saving and New Product Horizons

Beyond TVs, TCL’s application of small models has delivered measurable gains in areas such as air conditioning. The company’s compressors now leverage real-time optimization algorithms that significantly reduce energy consumption—reportedly by up to 40% in initial operating hours, with sustained savings exceeding 18% thereafter. This advancement not only lowers household costs but also supports broader environmental, social, and governance (ESG) objectives, reflecting a tangible example of AI-driven efficiency in mature product categories.

TCL has also begun channeling AI into entirely new product segments. Two notable directions in its roadmap include AR glasses—deeply integrated with large language model interactions—and a future generation of household robots designed as companion-oriented devices. These efforts signal a strategic move beyond incremental product enhancements toward what the company describes as “AI-native, AI-rich” innovations, where artificial intelligence fundamentally defines the user experience and product functionality.

Globally, TCL works closely with silicon and platform partners to enable on-device AI where possible—and falls back to cloud‑edge collaboration when needed. “Local computing is limited,” Sun acknowledged. “That’s why a hybrid architecture is essential for bringing advanced AI to everyday devices at a feasible cost.”

AI with Purpose

Throughout the interview and the factory visit, one theme remained constant: TCL’s AI strategy is deliberately vertical, practical, and consumer-centric. Rather than joining the race to build the largest foundation model, the company is focused on injecting its deep industrial knowledge into adaptable AI systems—systems that enhance today’s products, optimize its factories, and quietly enable the next generation of smart living.

As Daniel Sun put it: “The anchor should never be pursuing technology for its own sake. It must always return to consumers and customers—that’s where the real focus should be.”

For TCL, that focus is already showing—in quieter air conditioners, sharper displays, and smarter factories that together form a blueprint for AI done right.

Gizmochina attended TIC 2025 as an invited media partner. Special thanks to TCL for facilitating the interview and factory tour.

The post Exclusive: How TCL is Redefining AI with Vertical Integration and Advanced Manufacturing appeared first on Gizmochina.

Xiaomi has unveiled its most advanced open-source large language model to date, called MiMo-V2-Flash, as part of its expanding push into foundation AI. The new model focuses on high-speed performance and an efficient architecture, with strong capabilities in reasoning and code generation.

Xiaomi positions MiMo-V2-Flash as a direct competitor to leading models such as DeepSeek V3.2 and Claude 4.5 Sonnet. Let’s take a closer look at how the model works, its key features, and how to access it.

Purpose-Built for Speed and Agents

MiMo-V2-Flash is a Mixture-of-Experts (MoE) model with 309 billion total parameters and 15 billion active parameters. The model is purpose-built for AI agent scenarios and multi-turn interactions that require fast inference.

Xiaomi uses a 1:5 hybrid attention architecture, which combines Global Attention and Sliding Window Attention (SWA) with a 128-token window. The native context length is 32,000 tokens, and the model is trained with support for up to 256,000 tokens.

This design helps MiMo-V2-Flash maintain high efficiency while scaling across long-context tasks. Xiaomi claims it delivers output faster than several leading models, including DeepSeek and Claude, while maintaining lower operational costs.

Benchmark Performance and Pricing

Benchmark results show MiMo-V2-Flash performing at the top tier across various domains. The model ranks in the top two among open-source models in reasoning tasks such as AIME 2025 and GPQA-Diamond.

In software engineering benchmarks like SWE-Bench Verified and SWE-Bench Multilingual, it outperforms other open-source models and reaches levels comparable to GPT-5 and Claude 4.5 Sonnet.

Xiaomi has priced the API at $0.1 per million input tokens and $0.3 per million output tokens. The API is currently available for free for a limited time. According to the company, MiMo-V2-Flash generates responses at 150 tokens per second, while maintaining only 2.5% of Claude’s inference cost.

Technical Innovations Inside

The architecture includes Multi-Token Prediction (MTP), which allows the model to generate multiple tokens in parallel and verify them before output. This method increases decoding throughput without increasing attention or memory overhead. Xiaomi reports that with a three-layer MTP, the model reaches 2.0 to 2.6 times speed improvement compared to standard decoding.

Xiaomi also introduced a new post-training method called Multi-Teacher Online Policy Distillation (MOPD). The technique uses multiple teacher models to guide the student through token-level rewards in an on-policy learning process. It allows the model to achieve high performance with less than 1/50th of the training resources needed in traditional RL pipelines. MOPD also supports plug-and-play teachers, enabling continuous self-improvement cycles.

How to Access It?

Xiaomi has launched a web AI chat interface called MiMo Studio at aistudio.xiaomimimo.com, allowing users to interact directly with the model. The service supports web search, agent workflows, and code generation. It also features a toggle for switching between instant replies and slower “thinking” responses for deeper reasoning.

The model can generate functional HTML web pages and integrates well with development tools like Claude Code and Cursor. Xiaomi has also showcased creative and functional web demos.

Fully Open-Source

MiMo-V2-Flash is fully open-source under the MIT license. Model weights are available on Hugging Face, and all inference code is published on GitHub.

The company contributed inference code to SGLang on launch day and aims to grow developer adoption by offering transparent, low-cost access to high-performance AI tools.

MiMo-V2-Flash reflects Xiaomi’s shift toward becoming a serious player in the AI space. It brings competitive reasoning, fast code generation, and efficient agent deployment to the open-source ecosystem.

In related AI news, China has equipped traffic police with AI-powered smart glasses for real-time vehicle inspections, while a separate report highlights how even so-called “all-AI companies” still require human oversight due to limits in autonomous decision-making.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

The post Xiaomi MiMo-V2-Flash LLM Just Dropped: These Are the Most Interesting Things About It appeared first on Gizmochina.

Samsung has confirmed that its SmartThings AI Energy-Saving Mode can reduce the energy consumption of high-efficiency washing machines by about 30%. The data comes from a large-scale real-world study conducted in partnership with global carbon verification agency Carbon Trust.

The study analyzed energy usage data from approximately 187,000 Samsung washing machines in 126 countries. It ran for one year, from July 2024 to June 2025. The machines included in the study met high-efficiency standards in their respective markets, such as Energy Star in the US and 5 Star in India.

Samsung reported that enabling AI Energy-Saving Mode helped save a total of 5.02 gigawatt-hours (GWh) of electricity. This is equivalent to the monthly power consumption of around 14,000 households in Seoul, based on average figures from Korea Electric Power Corporation.

The AI Energy-Saving Mode uses SmartThings to monitor user behavior and optimize energy usage across compatible devices. The feature was activated manually by users during the study, which means the 30% savings figure came from voluntary usage rather than a controlled environment.

The verification followed the standards set by the Decarbonizing the Use-Phase of Connected Devices (DUCD) initiative. This initiative, led by Carbon Trust, provides energy consumption measurement guidelines for connected appliances. Samsung said it was the first company in the industry to carry out such large-scale verification based on DUCD protocols.

Samsung plans to expand its energy-saving efforts by developing more AI-driven appliances. The company said it will provide consumers with transparent data on energy use and carbon emissions to build trust and encourage adoption.

Samsung will showcase new AI-powered home appliances at CES 2026 during its ‘The First Look’ event in Las Vegas in January.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

(Source)

The post Samsung says its SmartThings AI cuts power use by 30% appeared first on Gizmochina.

Google thinks the future of the web isn’t just about making it agentic; it’s also about letting your browser build things for you. That idea is now real with “Disco”, the company’s experimental new AI-powered browser that vibe-codes an app based on what you are doing online. This feature of Disco is called “GenTab,” and is powered by Gemini 3.

As you might’ve guessed, GenTab takes your tab and chat history into consideration. It analyzes your tabs, your searches, and your chat prompts to generate tools that fit what you’re doing.

Google showed off a few examples, and they’re honestly pretty impressive. If you’re researching science topics like Entropy, Disco might spin up an “Entropy Explainer” app. The official video also shows a vibe-coded bunk bed comparison site and a memory match brain game.

Google’s new AI browsers vibe codes an app based on your tabs

GenTabs sit alongside regular tabs, but they get their own Gemini-like icon instead of a favicon. One demo showed a travel planner with calendars, route maps, crowd-level predictions, and quick-action buttons like “Book Nearby Stays.” Tap anything inside that app, and the GenTab reshapes itself in real time.

Disco greets you with a homepage containing a chatbox rather than the usual address bar. That’s where the chat history comes from, although you can also paste a URL in it.

If you end up conversing, Google will first suggest you relevant webpages, and after a few conversations, it pops up a prompt to create a GenTab based on what you’re searching.

AI is definitely changing the way we search and consume content. And it won’t surprise me that features like GenTabs are the next evolution of it. 

You don’t need to write code; you just describe what you need, refine it in plain language, and the browser does the heavy lifting. Google says Disco is meant to help people “learn faster” and experiment with what browsing could become. And yes, the company admits the best concepts from Disco might eventually show up in Chrome.

For now, Disco is an early experiment. Google Labs has opened a waitlist, and the first version is rolling out only on macOS. 

For more daily updates, please visit our News Section.

Don’t miss a thing! Join our Telegram community for instant updates and grab our free daily newsletter for the best tech stories!

The post Google’s new browser wants to “vibe-code” apps from whatever you’re doing online appeared first on Gizmochina.

Tiiny AI, a US-based startup, has announced the Tiiny AI Pocket Lab, which it claims is the world’s smallest personal AI supercomputer. The device weighs just 300 grams, fits in one hand, and delivers up to 190 TOPS of AI compute performance. It has been officially certified by Guinness World Records as the ‘Smallest MiniPC (100B LLM Locally)’ and can run up to 120-billion-parameter large language models (LLMs) entirely on-device.

The Pocket Lab features a 12-core ARMv9.2 CPU paired with a custom-designed NPU, achieving an AI compute performance of approximately 190 TOPS. It includes 80GB of LPDDR5X memory and a 1TB SSD to support large-scale inference workloads. Tiiny AI has built the device to operate within a 30W TDP and a typical 65W system power envelope, offering high performance with reduced energy consumption.

Tiiny AI says the Pocket Lab addresses the growing concerns around cloud dependency, data privacy, and rising energy costs. The company believes local AI processing offers a more sustainable and secure alternative to cloud-based systems. The device stores all user data and model interactions locally with bank-level encryption, eliminating the need for internet access or external servers.

The Pocket Lab supports one-click deployment of several open-source LLMs and agent frameworks. These include GPT-OSS, Llama, Mistral, DeepSeek, Qwen, and Phi, along with automation tools like ComfyUI, SillyTavern, and Flowise. The system runs entirely offline and enables use cases like multi-step reasoning, content generation, deep context understanding, and secure personal memory.

The company plans to showcase the Tiiny AI Pocket Lab at CES 2026. It has positioned the product for developers, researchers, professionals, and students who require portable and private AI capabilities. The device aims to serve real-world use cases in the 10B to 100B parameter range, which the company says accounts for over 80% of practical AI demands.

In related news, Xiaomi is reportedly developing a new AI assistant called Mi Chat, and reports indicate that GPUs are no longer the main barrier to AI progress, with other system-level constraints now becoming the limiting factor.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

(Source, Via)

The post Tiiny AI unveils a pocket-sized Mini PC that runs 120B LLMs locally appeared first on Gizmochina.

It feels like many products have got a new life with AI now, and that’s also the case for the earbuds category. These are the TicNote Pods 4G AI Recording Earbuds. Pay attention to the naming; you will know exactly why they have such a long name. Now these earbuds pack a ton of features. If you are a businessman, student, or journalist, you can really take advantage of them a lot. These earbuds are on Kickstarter time, and an important note to take is that, there are one or two features that I won’t be able to show you; they will only be available before the official launch.

Main Features 

Now what can these really do? Simple, AI Recording and transcription, 4G network on the product itself.

AI Recording

For the first time you use them, a very important note: remember to connect your phone, or let’s say your App, to both the case and the earbuds. Why? because the way to use them is different. If you are in a meeting, the case will be recording the audio. The same thing happens when you are in a one-on-one conversation, you use the case to record the other person’s talking. Either way, you will listen to the audio with the earbuds.

The other way to use the recording feature is with the earbuds. If you have an iPhone, even with the new iOS version, the other person still knows you are recording. If you have these, you can record the phone call with peace of mind. Or if you connect them to your computer, you can also record your ZOOM meetings.  

Now they can support a 25hrs of recording with the case, and 5 hours of phone recording with the earbuds. 25hrs of recording with the case is impressive, but it is kind of understandable, I mean, the case has a big battery we all know it. And also they have 32GB of storage built inside of it, so it can record for a long long time, while not taking too much of your phone’s storage.

The quality of the recording is also good, with ENC’s support, the voices can be recognized clearly.

AI translation

After the recording, you can get to the App to manage the files. To my experience, as long as the case is close to your phone, the files can be transferred from the case to the App fairly quickly. Ok so how is the transcription?

I have a one-hour-long meeting in French here, and I think it managed to do the job really well. It separates each speaker clearly, and records everything in the finest detail. It says there is voiceprint recognition tech inside, so it can identify each speaker, and then mark an ID on them in the App. I think it is super helpful, let’s say it’s a long boring meeting, and you just need to get the info and CTAs from your boss, then you just check what he or she said, and do what you have to do right away. One minor limitation in this part is that, it only supports Chinese and English, I hope more languages will be supported in the future.

The summary is very detailed and structured, so is the mind map. If you are like me who doesn’t know French, then the summary or the transcript can also be changed to English. The transcription speed is also very good. For this 1hr-ish audio, it managed to finish in like 3 mins.

The product supports 100+ languages transcription, which covers basically the languages from Mars. And tons of templates to be used for summary. You have 600mins of free transcription allowance every month, so make sure you use them for important meetings, otherwise you will have to buy some more.

Shadow AI

And then there is the so-called Shadow AI built in, basically it has almost all the advanced LLMs inside, including GPT-5, Claude, Gemini, and Grok. We know these AIs have their own advantages, having them all in one product is awesome, you get the best results whatever industry you are in, whatever project you are working on, and you save tons of money.

The Shadow AI helped with the work a lot. There are two parts in the App, the Aha moment and the Deep Research. The Aha moment is something worthy or creative you might have missed during the meeting. And the Deep Research can help with making some further solutions or ideas based on the conversation inside.

To save you the time of summarizing all the CTA and setting all the reminders, there is a feature called Intelligence Push. It fetches all the CTAs in that meeting, categorizes them, and sends a notification at the time you requested. This is a killer feature, it can really save you a lot of brain power and time to do this task.

You can also upload recordings or files, then ask Shadow AI to do the job. Let’s say you forgot the TicNote device and used your phone to record the meeting instead, or you have some old long files, you can just take advantage of the AI in this product.

TicNote Cloud

If you prefer to manage all the recordings, files on your computer, there is also the Ticnote Cloud. It is a webpage-based platform, where all the files are synced, and you can check and copy them. I find the efficiency of the cloud is very good. Right after I uploaded the recording to the App, the audio, the summary, all of them are on the webpage also.

Design

I won’t make you hear all the stuff about the design anymore, we have talked about earbuds quite enough. Basically they are well-made, if not better than all those open-earbuds out there. The shape, you must have seen similar ones. The materials, either flexible and also skin-friendly. Touch controls on the logo here. One thing to keep in mind though, there is a metal plate with the logo on the case. It can add some aesthetic to the product, but be careful not to put it in the same place as your phone or something delicate, or it will do a bit of damage to other things.

Other Features

Like I mentioned earlier, there are still some features not done yet, like the 4G network on the product itself, basically eSIM support. I don’t know how that would work, but from the info we received from the brand, it says a few years of free data allowance, and many countries are supported. You can refer to their website for the exact info.

Summary

So what do you think about the TicNote Pods 4G AI Recording Earbuds? Leave your comment below. After I tried all the features, I think they are an excellent tool for many people. Honestly, we have seen the AI features integrated in the earbuds this year, but none of them have gone far enough to support you with so many features. The super long recording time, internal storage, the very sophisticated AI features, and the upcoming 4G connectivity. I have never thought they could be so versatile and capable. So if you are a businessman who deals with tons of meetings, a student who would be buried by all kinds of notes, these are your best companions so far.

The post TicNote 4G AI Recording Earbuds: AI Transcription with 4G eSIM?! appeared first on Gizmochina.

OpenAI is moving beyond software and stepping directly into consumer hardware through a partnership between CEO Sam Altman and legendary designer Jony Ive. The company has already completed its first prototypes, marking a significant milestone in a project that has remained one of Silicon Valley’s most closely guarded secrets.

Here’s what the project looks like so far.

Rethinking Personal Tech

The new AI hardware concept is not meant to replace a phone or laptop. It is being developed as a ‘third device’ that sits alongside them. OpenAI CEO Sam Altman described the product as something that delivers a calmer experience, unlike the notification-heavy environment of modern smartphones. He compared current devices to walking through Times Square, while this new product aims to feel more like sitting in a peaceful cabin by a lake.

The device is expected to stay passive until needed and only interrupt when it has something useful or important to share. It is being built as an ambient assistant that remains aware of context and responds when appropriate. Altman said the goal is to give users a sense of peace and trust over time, without constantly demanding attention.

Screenless Device With Ambient Awareness

The hardware is screenless and likely pocket-sized. Early prototypes have been described as simple and tactile. Some reports say the form factor resembles an iPod Shuffle, with a clip-on design that may be worn around the neck or placed on a desk. It is not expected to look like a phone, smart glasses, or smartwatch.

Interaction will happen through microphones and possibly cameras or sensors. The device will rely on OpenAI’s latest models to understand spoken queries and real-world context. It will listen, observe, and remember.

Altman said it should be able to recall things like where the user last placed their keys or what book they looked at in a store. The goal is to create a device that feels intuitive, aware, and helpful across long stretches of time.

The $6.4 Billion Acquisition That Started It

OpenAI moved into hardware seriously after acquiring Jony Ive’s startup ‘io’ in May 2025. The $6.4 billion equity deal gave OpenAI access to a team focused on AI-specific hardware design. Ive, best known for designing the iPhone and iMac at Apple, took on the responsibility of shaping this new device. He now leads design work for both OpenAI and the former io team under his creative firm LoveFrom.

However, the ‘io’ branding became a legal issue. In June, Google Ventures-backed startup iyO sued OpenAI, claiming that the names ‘io’ and ‘iyO’ were phonetically similar and could confuse consumers. The Ninth Circuit Court agreed and barred OpenAI from using the name for any product similar to iyO’s. OpenAI removed all io branding from its website and public material.

Why Hardware Matters to OpenAI’s Future

OpenAI’s push into hardware has three main goals. First, it wants to collect real-world data through microphones and sensors, data that cannot be scraped from the internet.

Second, it wants more control over how its models reach users, without relying on Apple, Microsoft, or Google devices. Third, Altman wants to build a product that fulfills the long-term vision of an AI companion, similar to what was shown in the movie Her.

The company has described the device as a natural extension of its software. It will work with existing phones and computers, but operate in a separate category of its own. Altman said this is not about adding a screen or another app but creating a fundamentally new type of computing device.

What the Timeline Looks Like

The device is currently in prototype form. Altman and Ive confirmed in November 2025 that they have a working version, which they described as ‘jaw-droppingly good.’ Earlier versions did not meet their standards, but the latest iteration reportedly impressed both leaders.

Ive said the device could launch in under two years. Internally, OpenAI has targeted a release timeline of late 2026 or 2027. The company hopes to scale production quickly and ship 100 million units faster than any company has done before.

However, delays related to technical design, privacy concerns, and the ongoing trademark dispute could push the timeline further.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

The post Everything We Know About OpenAI’s First Hardware Device appeared first on Gizmochina.

Xiaomi is reportedly developing a new intelligent Q&A assistant named “Mi Chat,” as detailed in a recent post by the WeChat public account “Du Jia.” The leak suggests that the initial version of the product has already been completed.

This consumer-facing application is expected to demonstrate Xiaomi’s increasing proficiency in artificial intelligence and is anticipated to be powered by the company’s proprietary large language model, MiMo-7B-RL, which was launched earlier this year as part of its AI expansion strategy.

Despite having only 7 billion parameters, the MiMo model previously surpassed larger competitors like OpenAI’s o1-mini and Alibaba’s Qwen QwQ-32B-Preview in key benchmarks, including AIME 24/25 and LiveCodeBench v5.

Xiaomi is expected to scale the model further, and Mi Chat will likely act as its front-end interface, allowing the company to bring advanced reasoning features into everyday consumer scenarios. This initiative supports the company’s wider strategy to merge AI with its ‘Human x Car x Home’ ecosystem.

Xiaomi Partner and President Lu Weibing confirmed that the company has been making substantial investments in AI for several quarters, noting that large-model development progress has surpassed expectations. He stated that Xiaomi’s next strategic objective is the deep integration of AI with the physical world.

Furthermore, Xiaomi founder Lei Jun is reportedly taking a direct role in strengthening the AI team. Luo Fuli, a significant recruit from DeepSeek who joined the MiMo division this year, is listed as a corresponding author on a recent paper published in collaboration with Peking University.

More details about Mi Chat are expected to be unveiled by Lu Weibing and Luo Fuli at Xiaomi’s 2025 Human Car Home Ecosystem Partner Conference on December 17 in Beijing.

You can also read our story on how GPUs are no longer holding back AI and why a new bottleneck is taking their place.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

(Via)

The post Xiaomi is reportedly developing a new AI assistant called Mi Chat appeared first on Gizmochina.

Key Highlights:

• ByteDance’s new agentic AI smartphone faces major pushback from major Chinese apps over security and fairness concerns.
  
• The company has restricted its AI agent, Doubao, after reports of blocked logins, account freezes, and privacy worries.
  
• The launch reveals both the promise and risks of fully autonomous AI systems that can control a device like a human.
  

ByteDance’s Agentic AI Smartphone Faces Major Pushback Across China

ByteDance’s ambitious new AI smartphone, the Nubia M153, is drawing significant resistance from major Chinese platforms, prompting the company to roll back several of its most advanced features. Powered by Doubao, an agentic AI capable of operating the phone like a human, this device was meant to showcase the future of autonomous mobile computing. Instead, it has triggered a wave of concerns over security, fairness, and user privacy.

Major Chinese Apps Push Back

Leading platforms, including Alipay, Pinduoduo, and Taobao, have restricted or blocked Doubao’s automated operations. These companies cite issues such as unfair engagement, potential misuse, and risks to financial security. Some users reported that their accounts were frozen shortly after the AI performed tasks inside sensitive apps, even when they logged in manually.

Only a handful of major apps, including Meituan and ByteDance’s own Douyin, continue to work normally with the AI assistant.

ByteDance Restricts Doubao’s Powers

In response, ByteDance has scaled back Doubao’s capabilities. The AI can no longer interact with financial apps, claim user incentives, or operate competitive games where automated play could offer unfair advantages. The most significant restriction concerns WeChat: users experienced app crashes and suspensions, leading ByteDance to temporarily disable Doubao’s ability to control it.

The company says it is now working with developers to set clearer rules to avoid blanket bans.

Privacy Concerns Add to User Hesitation

ByteDance insists that the phone does not store screen data on its servers and that users’ activity is not used to train AI models. Still, many early adopters remain cautious. Doubao’s “global memory,” which allows it to remember tasks across apps, has heightened worries about device-level surveillance and unintended data exposure.

Real-World Challenges

The Nubia M153 made headlines after viral demos showed Doubao navigating apps visually, scrolling, typing, posting content, and completing multi-step tasks without step-by-step commands. The system’s full-stack control surpassed traditional assistants like Siri or Google Assistant, which rely on limited API calls.

But the same power that excited AI enthusiasts also alarmed regulators, app developers, and ordinary users.

Future Plans and Industry Impact

ByteDance stresses that the Nubia M153 is a “technology preview” priced at 3,499 yuan, mainly for industry professionals. The company plans to license Doubao to more smartphone brands after refining safety controls. The episode highlights a larger industry shift as tech companies worldwide, like Apple, Google, and OpenAI, race to build action-capable AI assistants.

For now, ByteDance’s phone demonstrates both the potential and the pitfalls of giving AI near-human control over a device. As the debate continues, the industry faces a defining question: how much autonomy should AI really have?

Read More:

• Leaked Video Exposes Tesla’s Robot Secret: Was Optimus Being Remote-Controlled?
• Say Goodbye to Driving: Rivian CEO Says Cars Will Soon Run Errands Without You
• Xiaomi’s AI Gamble Pays Off Big: Nobody Expected Returns This Fast

(via)

The post ByteDance’s Agentic AI Phone Sparks Panic and Nationwide App Bans appeared first on Gizmochina.

Meta and Ray-Ban have launched the second generation of their AI-powered smart glasses in India. The Ray-Ban Meta Gen 2 glasses are now available through Ray-Ban India and authorized retailers, with prices starting at ₹39,900 ($444) and going up to ₹45,700 ($508) depending on the frame and lens combination.

Specifications

The new model features a 12MP ultra-wide camera capable of recording 3K Ultra HD video for up to three minutes. Meta has improved image quality with ultrawide HDR and plans to roll out new camera modes like slow motion and hyperlapse through future software updates.

Meta has also upgraded the audio system. The Gen 2 glasses use a five-micro speaker setup that offers louder sound, deeper bass, and better directional audio. The touchpad and capture button support tap-based controls for music playback, photo capture, and messaging.

Battery life now lasts up to 8 hours on a single charge. The included charging case provides an additional 48 hours of power. Fast charging allows users to reach 50% charge in just 20 minutes. Meta says the glasses also support up to 5.4 hours of voice calls and 5 hours of music playback.

The Meta AI assistant now supports full Hindi voice functionality, which users can initiate by saying “Hey Meta.” This allows the assistant to handle commands for actions like snapping photos, dispatching messages, and regulating music. The assistant can now speak using Deepika Padukone’s AI-generated voice, which joins Meta’s global lineup of celebrity voices.

Meta is also set to introduce support for UPI Lite payments in India. This feature will allow users to scan QR codes using the glasses and finalize transactions through WhatsApp-linked bank accounts.

The Ray-Ban Meta Gen 2 collection includes updated versions of the Wayfarer, Headliner, and Skyler styles. Seasonal colors such as Shiny Cosmic Blue and Mystic Violet are also available.

In related news, we recently covered a detailed guide on what to look for when buying smart glasses, so make sure to check that out as well.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

The post Ray-Ban Meta (Gen 2) AI Glasses launched in India with 3K video, Meta AI in Hindi & UPI support appeared first on Gizmochina.

Meta has introduced a new research-stage tool called WorldGen that can create fully interactive 3D environments from a single line of text. The system can generate immersive, navigable virtual scenes in around five minutes. Meta said the content can be exported directly to popular game engines like Unity and Unreal without any conversion.

WorldGen combines several AI techniques, including procedural logic, diffusion-based 3D generation, and object-oriented scene decomposition. The output includes not only visual elements but also a complete navigation mesh, allowing characters to move through the environment without issues such as blocked paths or missing physics.

The system uses a four-stage pipeline. It starts with a layout plan created by a large language model. Then it reconstructs the scene geometry while ensuring walkable areas stay unobstructed. In the third step, it breaks down the scene into individual objects using a method called AutoPartGen. Finally, it adds high-resolution textures and final geometry polish to improve visual quality up close.

Meta said WorldGen solves common problems with earlier AI-generated 3D scenes. Many of those scenes focus on visual fidelity but lack collision data or physics support. WorldGen outputs mesh-based scenes with built-in functionality for simulation and gameplay use.

The current version supports environments up to 50×50 meters. Meta plans to expand this in future updates. The system does not reuse objects yet, which may affect performance in larger worlds. Despite that, Meta believes WorldGen will help teams save time in prototyping stages.

The company emphasized that WorldGen is still in development and not available to developers yet. However, it sees potential for the system to assist in game development, training simulations, and digital twin design by reducing the manual labor needed to create usable 3D environments.

In related AI news, a Japanese man who used AI to recreate a copyrighted image now faces criminal charges, and an AI-powered teddy for children was shut down by OpenAI after it delivered harmful responses.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

(Via)

The post This new AI by Meta creates game-ready 3D worlds in minutes appeared first on Gizmochina.

Google has officially introduced Gemini 3, its most advanced AI model to date. This new release builds on the foundation laid by Gemini 1 and 2, adding major upgrades in reasoning, multimodal understanding, and agent-based coding. Gemini 3 now powers several Google products, including Search, the Gemini app, and the company’s new developer platform, Google Antigravity.

Smarter, More Honest Answers

Gemini 3 marks a shift in how Google wants its AI to interact with users. The model avoids pleasantries and focuses on delivering accurate, clear, and helpful answers. Google says Gemini 3 gives smart, direct responses that focus on real insights instead of repeating user expectations.

This update reflects Google’s broader goal of turning AI into a thought partner. Gemini 3 understands subtle context and adapts its tone based on the task. Google claims this makes the AI more honest and productive during real-world use.

Better at Understanding Everything

Gemini 3 supports native multimodality. The system can handle text, images, video, audio, and code in one workflow. Users can upload mixed content such as handwritten notes, screenshots, or full lectures, and the model can interpret everything together. It can then generate interactive flashcards, concise summaries, or visual simulations based on the material

Google has increased the model’s context window to one million tokens. This lets Gemini 3 read and retain extremely long documents and conversations, improving performance on research-heavy or academic tasks.

Leading All the Benchmarks

Gemini 3 outperforms other frontier models across nearly all major AI benchmarks. It currently ranks first on the LMArena leaderboard with an Elo score of 1501. The model scores 91.9% on GPQA Diamond, 37.5% on Humanity’s Last Exam, and 23.4% on MathArena Apex.

For multimodal tests, Gemini 3 achieves 81% on MMMU-Pro and 87.6% on Video-MMMU. Its factual accuracy reaches 72.1% on SimpleQA Verified. These results show Gemini 3’s strength in both academic reasoning and real-world problem solving.

For developers, Gemini 3 leads the WebDev Arena with a score of 1487. It achieves 76.2% on SWE-Bench Verified and 54.2% on Terminal-Bench 2.0. These benchmarks confirm its skill in software development, command-line tasks, and autonomous coding.

Agents That Can Code on Their Own

Google has launched a new agentic development platform called Antigravity. The platform is powered by Gemini 3 and allows agents to write, test, and execute code across a full development environment. Gemini agents can access the code editor, browser, and terminal directly.

The system uses Gemini 3’s reasoning and tool-use skills to handle full-stack software tasks. In internal tests, it planned and built a working flight tracker app from scratch. The agent validated its work using browser-based simulations without human guidance.

Interactive Results in Search

Gemini 3 is now available in Google Search’s AI Mode. It powers dynamic visual layouts, interactive tools, and simulations directly in search results. When users ask complex questions, Gemini 3 creates real-time responses that include tables, charts, or even custom calculators.

If someone searches for topics like orbital physics or financial comparisons, Gemini 3 can generate interactive modules to help explain the results. These are generated using its built-in code writing and UI design capabilities.

New Advances in Training

Google DeepMind researchers credit Gemini 3’s success to progress in both pre-training and post-training. VP Oriol Vinyals said the improvements from Gemini 2.5 to 3.0 were the largest the team had seen. He also said pre-training had not hit a limit and there was still room to grow.

Many experts believed models had exhausted the internet’s public data. Google appears to have found ways to extract more value from existing data or has used internal datasets unavailable to competitors. This gives Google a long-term advantage in building next-generation AI systems.

Stronger Safety and Testing

Google says Gemini 3 is its safest model yet. The system has been tested against prompt injections, flattery behaviors, and misuse risks. External security groups like Vaultis and Dreadnode took part in the evaluations. Google also partnered with organizations such as the UK AISI for responsible development reviews.

Gemini 3 Pro is now available globally through the Gemini app and Google Search’s AI Mode for Pro and Ultra subscribers. The more advanced “Deep Think” version is still under safety review and will launch later.

In related AI news, iFlytek has introduced its AI Translation Earbuds with an open-ear design, real-time interpreting, and up to forty-two hours of battery life.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

The post Gemini 3 Just Dropped: Here Are the Most Interesting Things You Should Know appeared first on Gizmochina.

iFlytek has launched its latest AI Translation Earbuds in China with a pre-sale price of 2,499 yuan ($352). The new earbuds feature an open ear-hook design and support real-time translation across calls, face-to-face conversations, and lectures. iFlytek has also integrated its AI assistant, which can summarize content and respond to questions directly on the device.

iFlytek AI Translation Earbuds Specifications

The earbuds use an air conduction speaker system paired with a bone conduction sensor. This setup captures voice inputs from both near-field and far-field environments. Each earbud includes a three-microphone array that filters background noise using iFlytek’s multi-sense fusion noise reduction system. The device aims to deliver clear voice pickup during meetings, calls, and in-person interactions.

The translation system uses iFlytek’s end-to-end AI simultaneous interpreting model and supports a vocabulary of more than 100,000 professional terms across sectors such as healthcare, finance, and manufacturing. According to the company, the Chinese-English interpretation engine has a delay of only two seconds for the first word and delivers translated voice output with over 90 percent voice fidelity.

Offline translation support will roll out in a system update planned for March 2026. The update will enable translation across 18 language pairs without requiring internet access. The earbuds can also record bilingual conversations, saving both the original and translated audio along with text transcriptions on the connected smartphone.

The AI assistant can summarize meeting notes, process article links, and generate voice podcasts in a two-person conversational style. These software features currently support Chinese and English.

Each earbud uses an 11 mm high-sensitivity driver supported by dynamic low-frequency compensation, which strengthens bass response typically lost in open-ear designs. The system includes DRC dynamic range control, which monitors sudden peaks and prevents the audio from clipping. This keeps loud passages stable and protects overall listening comfort. iFlytek uses skin-friendly materials and includes a silver-ion antibacterial coating for hygiene.

The earbuds use Bluetooth 6.0 to stay connected to two devices at the same time, such as a phone and a laptop. They deliver up to six hours of translation or twelve hours of music on a single charge, and the case increases total runtime to roughly forty-two hours. They also support fast charging, and a ten-minute top-up adds about three hours of playback. The earbuds carry an IP55 rating that protects them against sweat and light rain during everyday use.

In related news, JVC has debuted its Victor WOOD Master flagship wireless earphones featuring hybrid wooden drivers and LDAC support, while Dell has introduced the Pro Plus Earbuds in India with AI noise cancellation and adaptive ANC.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

(Source)

The post iFlytek launches AI Translation Earbuds with open-ear design, real-time interpreting & 42h battery appeared first on Gizmochina.

Anthropic has reported what it calls the first documented large-scale cyberattack executed primarily by an AI system. The company said the incident occurred in mid-September 2025 and involved its Claude Code tool being misused to carry out a sophisticated, multi-stage intrusion campaign.

According to Anthropic, the attackers used agent-like AI capabilities to perform up to 90 percent of the work. Human involvement was minimal, limited to a few key decisions during each cycle. The company stated that hackers posed as cybersecurity researchers and used jailbreak techniques to bypass safeguards. They instructed Claude to complete a series of small tasks that, together, formed an automated cyber-espionage operation.

Anthropic said the attackers targeted nearly 30 organizations, including tech companies, financial institutions, chemical manufacturers, and government agencies. The AI system handled reconnaissance, wrote exploit code, extracted credentials, identified vulnerabilities, and exfiltrated sensitive data. It also compiled detailed reports that helped the attackers plan future operations.

The company said its investigation linked the attack to a Chinese state-sponsored group, although it did not disclose the specific evidence. The Chinese embassy denied involvement. Anthropic has banned the attackers from using Claude and notified affected organizations and law enforcement agencies.

Cybersecurity researchers have expressed skepticism. Martin Zugec from Bitdefender said the report lacked verifiable threat intelligence and called the claims speculative. He acknowledged the growing risk of AI-driven threats but stressed the need for transparency.

Anthropic argued that AI tools are essential for both attack and defense. It used Claude to help analyze the breach and called for stronger AI safeguards and better coordination among industry and government.

The report suggests that AI systems can now perform tasks that once required entire teams of expert hackers. Anthropic warned that the barrier to advanced cyberattacks has dropped and urged organizations to adopt AI-powered security tools to counter emerging threats.

In related news, reports suggest AI search engines increasingly surface low-traffic websites over major sources, and LSE predicts Gen Alpha could rely on voice rather than keyboards at work by 2028.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

(Source, Via)

The post Anthropic claims China-backed group used AI for massive cyberattack appeared first on Gizmochina.

The keyboard could soon become obsolete, at least that’s the future outlined by a new study from the London School of Economics and Jabra. According to the report, by 2028, voice-based AI will become the dominant interface for workplace communication.

For Generation Alpha, those born after 2010, the keyboard might never even be part of the daily workflow. Instead, they’ll speak to their devices first, edit later.

Paul Sephton, global head of brand communications at Jabra, told Fortune that in the AI-powered workplace of the near future, the first draft of work will be spoken, not typed. Typing will serve only as an editing step, not a creative one.

It’s a shift the report says mirrors how humans naturally think: fast, iterative, conversational. It also opens the door for more spontaneous creativity. Voice input offers a clear productivity edge for working parents and multitaskers, enabling hands-free interactions on the go.

But not everyone is buying into the hype. Experts warn that this voice-first future might introduce new problems. Management professor Fabrice Cavarretta from ESSEC Business School said that while speaking may become the dominant input method, it won’t replace written communication. Voice messages lack skim-ability, keyword search, and clarity. They are harder to archive and slower to process.

Bertrand Audrin from EHL Hospitality Business School added that unless voice notes are transcribed, accountability suffers. Voice lacks permanence and can become a burden in workplaces that rely on traceable decisions and clean communication logs. Editing raw speech into coherent text isn’t seamless either. It demands cleanup, especially for non-native speakers or informal tone corrections.

So while the desk-bound keyboard might be on the decline, it won’t vanish without a fight. The future may be spoken, but it will still be written, transcribed by AI from what we say.

In related AI news, Meizu has launched the StarV Snap AI glasses featuring the Snapdragon AR1 chip, a 12MP camera, and a built-in AI assistant. Meanwhile, new reports suggest that AI search engines are increasingly favoring low-traffic websites over major sources.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

(Source)

The post Gen Alpha will use voice instead of keyboards at work by 2028, predicts LSE appeared first on Gizmochina.

1X Technologies, a robotics startup backed by OpenAI, has opened preorders for its first consumer-facing humanoid robot named Neo. The robot costs $20,000, with shipments set to begin in the United States in 2026, followed by a global release in 2027.

Specifications and Design

Neo stands 1.68 meters tall and weighs around 30 kilograms. It uses tendon-driven actuators and features a soft body made of 3D lattice polymer. The robot has a low operating noise of 22 decibels and can operate for up to four hours on a single charge. It can carry loads up to 24.95 kilograms and lift objects weighing up to 69.85 kilograms. Neo is available in three color options: Tan, Grey, and Dark Brown.

1X says Neo can take care of everyday household chores, including folding laundry, organizing shelves, and cleaning rooms. The robot can be operated through voice commands or a dedicated mobile app that supports task scheduling and real-time monitoring. It also includes a built-in large language model that enables contextual conversations, answers user queries, and offers helpful suggestions.

Neo combines audio, visual, and memory-based data to personalize its behavior and recognize when it is being addressed. It can also process visual information to identify kitchen ingredients and suggest recipes accordingly.

During its early stages of development, Neo relies on a feature called Expert Mode for tasks it does not yet understand. Through this system, users can schedule remote assistance sessions where a human operator from 1X can take control using a VR headset. These sessions require user consent, and privacy options let users blur faces, define private zones, and restrict access.

1X says it has implemented multiple security layers to protect user privacy. The company emphasized that early adopters should expect some level of human teleoperation to support Neo’s AI training. CEO Bernt Øyvind Børnich described this as part of a “social contract” between users and the company to help advance the technology.

Pricing and Availability

The robot is now available for preorder on the official 1X website with a refundable deposit of $200. The company also offers a $499 monthly subscription plan for those who prefer not to pay the full price upfront.

In related AI news, a study finds that AI search engines tend to cite lesser-known websites over popular sources. Meanwhile, Google has introduced Vibe Coding in AI Studio, allowing developers to create AI apps from a single prompt.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

(Source, Via)

The post This $20000 humanoid robot will clean your house but there is a hidden catch appeared first on Gizmochina.

There’s a new study showing that AI-powered search engines often cite lesser-known websites more frequently than traditional ones like Google. Researchers from Ruhr University Bochum and the Max Planck Institute for Software Systems detailed the findings in their paper titled “Characterizing Web Search in the Age of Generative AI.”

The team compared Google’s standard search results with responses from Google’s AI Overviews, Gemini 2.5 Flash, and two variants of GPT-4o’s web search mode. The analysis revealed that AI systems often draw from websites that rank far lower in popularity. In many cases, these sources do not appear in Google’s top 100 results for the same query.

The researchers used a range of test queries, including real user questions from ChatGPT conversations, political topics from AllSides, and Amazon’s most-searched products. According to Tranco, a domain-ranking tool, the sources AI engines used ranked significantly lower than those in Google’s top results. Gemini’s results had a median domain rank beyond the top 1,000.

Over half of the sources cited by Google’s AI Overviews were absent from the top 10 traditional results. About 40 percent did not appear in the top 100 at all. However, this shift toward obscure websites did not appear to harm the quality of information. GPT-based models frequently cited corporate sites and encyclopedic content while avoiding social media.

The study found that AI-generated search results contained about the same range of identifiable “concepts” as traditional search, suggesting that AI systems maintain a comparable level of information diversity. However, because these models summarize information rather than listing individual sources, they often compress large amounts of data into shorter responses. This compression can lead to the loss of smaller or more ambiguous details that might still appear in traditional search results.

Researchers also noted that AI-powered search tools struggle with time-sensitive or rapidly changing topics. GPT-4o’s hybrid mode, for instance, sometimes failed to provide the most up-to-date information when responding to queries about recent events or trending subjects.

The future of search isn’t about better or worse. It’s about difference. And if this study is right, AI may be building a web that values synthesis over popularity, and that’s rewriting the rules of online authority.

In related AI news, Google has introduced Vibe Coding in AI Studio, a new feature that lets developers create fully functional AI apps from a single prompt. Meanwhile, ByteDance has launched Seed3D 1.0, a tool capable of instantly converting any 2D photo into a realistic 3D model.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

(Source, Via)

The post AI search engines prefer low-traffic websites over popular sources appeared first on Gizmochina.

ByteDance, the parent company of TikTok, just dropped a major piece of the future of 3D content creation. It’s called Seed3D 1.0, and it’s not your average model generator. This tool can turn a single 2D image into a complete, simulation-grade 3D model, complete with detailed geometry, photorealistic textures, and physically based rendering (PBR) materials.

It’s an end-to-end pipeline built on the increasingly popular Diffusion Transformer architecture, trained on massive datasets and clearly aiming for dominance in the generative 3D space.

Seed3D 1.0 focuses on realism, structure, and scalability

While tools like Gaussian Splatting and open-source projects such as Hunyuan3D have pushed boundaries, ByteDance claims Seed3D 1.0 outperforms both open-source and closed-source rivals in texture quality and geometric fidelity. It achieves this with just 1.5 billion parameters, beating even larger models like Hunyuan3D 2.1, which uses 3 billion parameters.

The core innovation lies in its combination of a multimodal Diffusion Transformer and a stepwise generation strategy. It begins by analyzing an image with a vision-language model to extract object-level and spatial cues. It then generates individual 3D models and assembles them into a full scene. This allows it to scale from a single chair to a fully detailed office or even a large-scale cityscape.

Seed3D 1.0 also excels at maintaining texture consistency across multiple viewpoints. Instead of relying on generic textures, it generates view-aligned materials that remain coherent from every angle, ensuring both realism and structural accuracy for simulation-grade use. The output isn’t just for research purposes; models created with Seed3D can be directly integrated into simulation platforms like Isaac Sim for embodied AI training.

As the boundary between real and synthetic content continues to blur, Seed3D 1.0 marks a significant leap forward for ByteDance, and potentially for 3D content creators, robotics developers, and spatial computing platforms around the world.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

(Source, Via)

The post ByteDance Unveils Seed3D 1.0: Turn Any Photo Into a Realistic 3D Model Instantly appeared first on Gizmochina.

Goldman Sachs has dismissed fears of an AI-driven market bubble, stating that the artificial intelligence boom is still in its early stages. In a new report released on Wednesday, analysts at the investment bank said current spending on AI remains modest when compared to the long-term economic potential of the technology.

The report estimates that U.S. AI-related investments account for less than 1% of GDP. In contrast, previous technology booms such as railroads, electrification, and the internet era peaked at 2% to 5% of GDP. Goldman expects AI investment to rise to $300 billion by 2025, citing continued improvements in productivity and expanding enterprise adoption.

The firm outlined two main reasons behind its optimistic outlook. First, AI deployments across existing sectors are already producing measurable productivity gains. Second, these gains depend on large-scale compute infrastructure, which justifies current investments in chips, servers, and data centers.

According to Goldman Sachs, generative AI could eventually add up to $20 trillion to the U.S. economy, with $8 trillion flowing to businesses as capital income. The bank projects that labor productivity could rise by 15% over the next decade once AI tools become widely adopted.

However, the report also pointed out a potential risk. It remains uncertain whether the companies investing most heavily in AI today will benefit the most in the long run. Past infrastructure cycles show that early entrants often built costly systems that were later acquired cheaply by more successful latecomers. The rapid depreciation of AI hardware and the fast pace of model development could make similar outcomes likely.

Despite that, Goldman believes the current environment still favors further AI investment. The firm expects spending to stabilize over time as the industry moves beyond the infrastructure build-out phase and hardware prices decline.

In related AI news, Google has recently introduced the Gemini 2.5 Computer Use model, featuring human-like web browsing and real-time interaction capabilities. Meanwhile, Alibaba has unveiled a massive 1-trillion-parameter AI model designed to rival ChatGPT and Google Gemini in scale and performance.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

(Via)

The post Goldman Sachs Says AI Hype Is Just Warming Up, Not a Bubble appeared first on Gizmochina.

Google has released a new AI model called Gemini 2.5 Computer Use. The model allows AI agents to interact with websites and user interfaces the way a human would. It is now available in public preview via the Gemini API on Google AI Studio and Vertex AI.

The model builds on Gemini 2.5 Pro’s visual understanding and reasoning capabilities. It can perform a wide range of browser-based actions such as clicking, typing, scrolling, hovering, opening dropdowns, and navigating through URLs. Google says the model outperforms competing tools on several benchmarks, including Online-Mind2Web, WebVoyager, and AndroidWorld, while maintaining lower latency.

Unlike traditional AI models that rely on APIs, Gemini 2.5 Computer Use processes screenshots of web interfaces and generates specific UI actions in response. The agent receives a task prompt, a screenshot of the digital environment, and a history of recent actions. It then analyzes the interface and returns a UI action, such as clicking a button or typing into a field. The action is executed on the client side, and a new screenshot is sent back to the model to continue the task in a loop.

Google demonstrated the model’s performance with examples that show the agent sorting sticky notes on a digital whiteboard and transferring pet details from one website to a CRM system. The demo videos are accelerated to show the process in real time.

The model supports 13 actions currently and works best with web browsers. Google said it is not yet optimized for desktop OS-level tasks, though it has shown potential on mobile benchmarks.

Google has also implemented safety measures to prevent misuse. Each action proposed by the model is reviewed by a safety service before execution. Developers can restrict certain actions or require explicit user confirmation for high-risk tasks like financial transactions.

Several internal Google teams are already using the model in production. It supports UI testing and automation tasks across platforms such as Search and Firebase. External developers in the early access program have used the model to build workflow automation and assistant tools.

Developers can start using the model through Google AI Studio or Vertex AI. Google also provides a demo environment via Browserbase for testing and experimentation.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

(Source)

The post Google launches Gemini 2.5 Computer Use model with human-like web browsing capabilities appeared first on Gizmochina.

So, Alibaba just released something that’s got the AI world talking. Meet Qwen3-Max, their latest and greatest language model that’s basically saying “Hey OpenAI, Google, and Anthropic, we’re here to play too.”

The Trillion Parameter Leap

Here’s the thing that caught everyone’s attention: this model has over 1 trillion parameters and was trained on a staggering 36 trillion tokens. That’s a massive jump that puts them squarely against GPT-5, Gemini 2.5 Pro, and Claude Opus 4. But here’s what’s really cool, they didn’t just make it bigger for bragging rights.

They used Mixture-of-Experts (MoE) architecture, which is like having a really smart team where only the right experts jump in when needed. Instead of activating the entire trillion-parameter network every time, it only fires up specific subsets during each forward pass. It’s way more efficient than just throwing raw compute at the problem.

But here’s where the engineering gets impressive: they implemented something called “global-batch load balancing loss” to keep training stable. The result? Their loss curve stayed smooth throughout the entire training process, no spikes, no need to restart, no mid-training data shuffling. Anyone who’s trained large models knows how rare that is.

Efficiency Breakthroughs

Now this is where things get really technical, and honestly, pretty impressive:

Training Speed: They developed PAI-FlashMoE, which is basically a multi-stage parallel pipeline optimization strategy. The result? 30% better training throughput compared to Qwen2.5-Max-Base. That’s not just a small improvement, that’s the difference between months and weeks of training time.

Long Context Magic: For handling really long documents, they created something called ChunkFlow strategy. This thing delivers 3x faster throughput than traditional sequence parallelism when dealing with long contexts. We’re talking about a 1 million-token context window here, that’s in the same league as Claude Opus 3 and Gemini 1.5 Pro.

Reliability Engineering: They built systems called SanityCheck and EasyCheckpoint that reduced hardware failure downtime to one-fifth of what they experienced with Qwen2.5-Max. When you’re running massive compute clusters, this kind of reliability engineering is absolutely crucial.

Performance

Qwen3-Max-Instruct (the version you can actually use) is putting up some seriously impressive numbers:

LMArena Leaderboard: Ranked #3 overall, and get this, it’s ahead of GPT-5-Chat. That’s genuinely impressive performance.

Real-World Coding: Scored 69.6 on SWE-Bench Verified, which tests whether AI can actually solve genuine GitHub issues and programming problems. That beats DeepSeek V3.1 and is competitive with Claude Opus 4.

Tool Usage: Hit 74.8 on Tau2-Bench, which measures how well models can actually use APIs and external tools. This beat both Claude Opus 4 and DeepSeek V3.1, which is honestly surprising.

But here’s where it gets really wild: they’re working on Qwen3-Max-Thinking, a reasoning-focused variant that’s still in training. In early tests that included tool usage and parallel compute during inference, it scored perfect 100% on both AIME 25 and HMMT, two of the most brutal mathematical reasoning benchmarks in AI. This version comes with an integrated code interpreter and focuses on complex logical problem-solving.

Multilingual and Multimodal Story

While the current public release focuses heavily on reasoning and coding, Alibaba built Qwen3-Max to handle multilingual tasks with particular strength in English and Chinese. The improvements are significant across instruction-following, mathematical reasoning, scientific tasks, and, this is important, it produces fewer hallucinations than previous versions.

The model shows major improvements in areas that actually matter for real-world deployment: better logic, more accurate math, stronger scientific reasoning, and more reliable responses overall.

How to Actually Use This Thing

For Regular Users: Qwen3-Max-Instruct is available right now through the Qwen app (iOS/Android) and their website. The app defaults to using Qwen3-Max, but you can manually switch if you want to compare with other versions.

For Developers: You can access it via API through Alibaba Cloud’s Model Studio, which means you can integrate it directly into your applications, tools, or services.

The Bigger Picture

Here’s what’s really significant: Alibaba isn’t just trying to match existing models, they’re pushing the technical frontier in specific areas like long-context processing, training efficiency, and system reliability. The engineering details show genuine innovation, not just scaling up existing approaches.

The upcoming Qwen3-Max-Thinking variant, with its focus on reasoning and tool integration, suggests they’re aiming to build something that can function as a genuine autonomous agent, not just a better chatbot.

Should You Care?

If you’re a casual user, this gives you another top-tier model to experiment with, especially if you need long-context processing or multilingual capabilities.

If you’re a developer, the API access and technical capabilities make it worth serious evaluation against your current toolchain.

If you’re interested in AI progress generally, this represents a real milestone: we now have multiple organizations capable of building and deploying trillion-parameter models with genuinely different technical approaches and capabilities.

The AI landscape just got a lot more competitive, and frankly, a lot more interesting.

In related AI news, researchers developed an AI-designed virus that targets bacteria, raising safety concerns, and Huawei unveiled its powerful Atlas 950 and 960 SuperPoDs to challenge Nvidia.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

The post Alibaba Just Dropped a 1 Trillion Parameter AI Model to Take On ChatGPT and Gemini appeared first on Gizmochina.

Picture this: a computer sitting in a lab, essentially “thinking up” brand new forms of life from scratch. That’s exactly what just happened at Stanford and the Arc Institute, where researchers used artificial intelligence to design viruses that can actually kill bacteria. And yes, these AI-created viruses work in real life.

Here’s how wild this is, it’s the first time anyone has used AI to dream up an entire viral genome from nothing. No copying nature’s homework, no tweaking existing designs. Pure digital creativity turned into living biology.

The AI system, called Evo, works kind of like ChatGPT, except instead of being trained on books and articles, it learned from 2 million different virus genomes. When researchers asked it to design new versions of a simple virus called phiX174, Evo created 302 completely original genetic blueprints. The kicker? Sixteen of them actually came to life in the lab and successfully infected E. coli bacteria.

“This was like watching digital code become biology,” said Brian Hie, who runs the lab. You can practically hear the excitement, and maybe a little nervousness, in his voice.

The potential upside is huge. We’re talking about supercharged treatments for antibiotic-resistant infections, which kill hundreds of thousands of people every year. Imagine having an AI design custom viruses to hunt down specific bacteria that antibiotics can’t touch anymore. Or think about gene therapy, using these designer viruses as tiny delivery trucks to fix genetic diseases.

But here’s where things get uncomfortable. J. Craig Venter, basically the godfather of synthetic biology, didn’t sugarcoat his concerns. He called this “a faster version of trial and error” and warned that the same technology could potentially be misused to create much nastier viruses than simple bacteria killers.

Right now, Evo only knows about viruses that don’t harm humans. But the underlying technology? That could theoretically be retrained on more dangerous stuff. It’s like having a recipe generator that currently only knows how to make cookies, but could learn to make dynamite if someone fed it the wrong cookbook.

We’re still years away from AI creating entire synthetic cells, those would need millions of genetic letters instead of the thousands in these simple viruses. But companies like Ginkgo Bioworks are already working on fully automated systems that could go from AI design to living organism without human hands touching the process.

It’s honestly breathtaking and unsettling at the same time. We might be witnessing the moment when life itself becomes programmable, when the line between digital code and biological reality finally disappears. The question isn’t whether this technology will change everything, it’s whether we’re ready for what comes next.

In related AI news, Huawei has recently unveiled its powerful Atlas 950 and 960 SuperPoDs in a direct challenge to Nvidia, while Tencent has introduced a groundbreaking 3D AI tool that is available completely free.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

(Via)

The post AI just created a virus that kills bacteria—should we be scared? appeared first on Gizmochina.

Huawei just made a massive power play in the AI hardware wars. At Connect 2025 in Shanghai, rotating chairman Eric Xu unveiled the Atlas 950 and Atlas 960 SuperPoDs – absolutely enormous clusters of Ascend accelerators that are clearly designed to challenge Nvidia’s upcoming Rubin-era systems (although Nvidia AI chips are banned now in China) and push China’s AI capabilities to new heights.

Atlas 950 SuperPoD

The Atlas 950 SuperPoD is built around 8,192 Ascend 950DT chips, and the performance figures are genuinely impressive. We’re looking at 8 EFLOPS of FP8 compute power and 16 EFLOPS at FP4 precision, backed by a massive 16.3 PB/s of interconnect bandwidth.

In practical terms, this translates to some serious AI workload performance: 4.91 million tokens per second for training and 19.6 million tokens per second for inference. That’s a 17× and 26× improvement over their current Atlas 900 A3 system, which is no slouch itself.

The physical footprint is equally massive – we’re talking 160 cabinets all connected via Huawei’s new UnifiedBus 2.0 optical protocol. Huawei claims this interconnect is ten times faster than today’s internet backbone infrastructure, which is critical when you’re moving data between thousands of accelerators.

SuperCluster

Here’s where Huawei’s ambitions become clear. They’re planning to link 64 of these SuperPods into Atlas 950 SuperClusters, creating systems with more than 520,000 NPUs delivering 524 EFLOPS of FP8 compute.

Xu wasn’t shy about the competitive positioning either; he directly claimed this setup will outperform Elon Musk’s xAI Colossus and even Nvidia’s upcoming NVL144 and NVL576 deployments, citing six to seven-fold compute advantages plus superior memory and network throughput.

Atlas 960

But Huawei isn’t stopping there. The Atlas 960 SuperPoD, slated for 2026, essentially doubles every major specification: 15,488 Ascend 960 chips, up to 30 EFLOPS FP8, 60 EFLOPS FP4, and 34 PB/s bandwidth. The companion 960 SuperCluster is projected to hit 2 ZFLOPS of FP8 performance when it arrives in 2027, numbers that start to sound almost surreal.

New Ascend Roadmap

Behind all this hardware is a completely refreshed Ascend chip roadmap. The 950PR and 950DT arrive in 2026, featuring Huawei’s own HBM variants: HiBL 1.0 for cost-efficient prefill operations and HiZQ 2.0 optimized for decode and training workloads.

The 960 follows in 2027, then the 970 in 2028. Each generation promises to “double compute” while expanding support for emerging formats like FP8, MXFP4, and HiF4.

Scale-Over-Silicon Strategy

For a company that’s been cut off from advanced Western semiconductor fabs, Huawei’s strategy is becoming crystal clear: if you can’t win on per-chip performance, win through sheer scale and system integration.

By controlling the entire stack, memory subsystems, networking fabric, packaging, and interconnects, and by deploying hundreds of thousands of chips behind ultra-low-latency interconnects, they’re betting they can meet China’s massive demand for model training compute while seriously challenging Nvidia’s market position.

Reality Check

Of course, there’s always a gap between conference announcements and real-world deployments. These massive systems will face significant challenges around power consumption, thermal management, and perhaps most critically, software ecosystem maturity. Nvidia’s CUDA advantage didn’t happen overnight, and replicating that software stack optimization across such massive clusters is no small feat.

But from a pure ambition standpoint, Huawei has definitely fired a serious shot across the Pacific. Whether these enormous machines can deliver their promised performance outside of PowerPoint presentations will largely depend on solving real-world engineering challenges around power budgets, cooling infrastructure, and software optimization.

The AI hardware race just got a lot more interesting.

In related AI news, Tencent has recently launched a groundbreaking 3D AI modeling tool that’s completely free, while Baidu’s new PP-OCRv5 model outperforms larger rivals in OCR benchmarks.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

The post Huawei Just Declared War on Nvidia With Its Monster Atlas 950 & 960 SuperPoDs appeared first on Gizmochina.

Remember when creating realistic 3D models meant weeks of painstaking work by skilled artists? Well, Tencent just threw that playbook out the window with the latest 3.0 version of Hunyuan 3D, and honestly, the results are pretty mind-blowing.

What Makes This Different?

Most AI 3D generators feel like they’re throwing digital clay at a wall and hoping something decent sticks. Hunyuan 3D 3.0 takes a completely different approach with something called “3D-DiT”, think of it as having a master sculptor who starts rough and gradually adds finer and finer details.

Instead of trying to generate everything at once (which usually leads to wonky-looking results), this system builds models layer by layer. It starts with basic shapes, then keeps refining until you get something that actually looks… well, real.

The Numbers Are Ridiculous

We’re talking about three times better precision than before, support for up to 36 billion voxels (that’s a lot of tiny 3D pixels), and resolution that can go up to 1536³. But here’s the kicker: it’s completely free to use.

Finally, Characters That Don’t Look Like Aliens

Anyone who’s played around with AI-generated 3D models knows the struggle: faces that look slightly off, bodies with weird proportions, poses that scream “I was made by a robot.” Tencent seems to have cracked this code.

The new system generates human characters that actually look human. Facial features sit where they’re supposed to, skin textures look believable, and poses flow naturally instead of looking like mannequins arranged by someone who’s never seen a person move.

Smart Problem-Solving

Here’s where it gets really clever: the system can figure out what should be there even when it can’t see it. If you’re generating a character who’s partially turned away or behind an object, the AI makes educated guesses about the hidden parts based on what it does see. No more bizarre gaps or impossible anatomy.

And textures? They actually wrap around objects properly now. No more skin that looks painted on or fabric that defies physics.

Built for Everyone (Seriously)

Tencent could have easily charged premium prices for this technology. Instead, they’re giving it away free through their Hunyuan 3D platform. For developers who want to integrate it into their own projects, there’s a cloud API available too.

If you’re a professional who needs more advanced features, they’re also testing Hunyuan 3D Studio, a full production suite that handles everything from UV mapping to rigging. It’s invite-only for now, but you can apply through their platform.

The Open-Source Promise

In a move that’ll make developers everywhere happy, Tencent plans to open-source a version called Hunyuan 3D Omni. This will give researchers and companies even more control over the generation process, perfect for when you need something very specific rather than just “make me a cool character.”

Why This Matters

Look, we’ve seen plenty of AI tools promise to revolutionize creative work, and most fall short when you actually try to use them. But Hunyuan 3D 3.0 feels different. The combination of actually good results, zero cost, and accessibility could genuinely democratize 3D modeling.

Whether you’re an indie game developer working on a shoestring budget, a digital artist exploring new possibilities, or just someone who thinks creating 3D stuff sounds fun, the barrier to entry just got obliterated.

Sometimes the most exciting tech advances aren’t the flashiest ones; they’re the ones that just work really, really well.

In related AI news, Baidu has recently released PP-OCRv5, a compact optical character recognition model that surpasses larger models in benchmark tests. The company has also unveiled the Ernie X1.1 deep thinking model, which it claims outperforms DeepSeek R1 in reasoning and comprehension tasks.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

(Via)

The post Tencent Just Dropped a Game-Changing 3D AI Tool And It’s Completely Free appeared first on Gizmochina.

Baidu just dropped something pretty interesting in the AI scene. After their recent launch of Ernie X1.1 deep thinking model, they’ve now released PP-OCRv5, a new optical character recognition model that’s available on Hugging Face. What makes this one stand out? It’s designed to be really good at reading text while staying surprisingly lightweight.

The thing is, those massive vision-language models we keep hearing about? They’re impressive, but they can struggle when it comes to the nitty-gritty work of reading structured text accurately. That’s where PP-OCRv5 comes in. Baidu built this one specifically to tackle those limitations head-on.

Here’s what’s cool about it: the model works in two main stages – first it finds where text is located in an image, then it actually reads what that text says. This approach helps it nail down exactly where text appears and draw precise boxes around it, which is super handy if you’re trying to pull data from documents or analyze forms.

The efficiency is pretty remarkable too. We’re talking about just 0.07 billion parameters – that’s tiny compared to the giants in this space. Baidu tested it on mobile setups and found it could churn through over 370 characters per second on an Intel Xeon processor. That means you could actually run this thing on regular computers or even edge devices without needing massive server farms.

When Baidu put PP-OCRv5 head-to-head with the big names like GPT-4o, Gemini 2.5 Pro, and Qwen2.5-VL on OCR tasks, their model came out ahead. It handles both printed and handwritten text pretty well, and it’s not just limited to English – it works with Simplified Chinese, Traditional Chinese, Japanese, Pinyin, and actually supports more than 40 languages total.

The technical setup is straightforward but smart. It starts by cleaning up the image – fixing rotation issues, reducing distortion, that sort of thing. Then it finds where text lines are located, figures out which way they’re oriented, and finally converts those characters into readable text. The whole process is designed to give you precise coordinates for where each piece of text sits, which is crucial if you’re scanning invoices or processing forms where layout matters.

What’s nice is that Baidu made this available to everyone through Hugging Face. For developers and businesses dealing with lots of multilingual documents or just needing solid OCR capabilities without the overhead of massive models, PP-OCRv5 looks like it could be a practical choice that actually gets the job done.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

(Via)

The post Baidu releases PP-OCRv5, a compact AI model that beats large rivals in OCR tests appeared first on Gizmochina.

Baidu is back with another AI announcement, and this time they’re really swinging for the fences. The Chinese tech giant just rolled out Ernie X1.1, which they’re calling a major upgrade to their previous Ernie X1 model.

What’s New Under the Hood?

Wang Haifeng, Baidu’s Chief Technology Officer, presented the new model as something special. Built on their Ernie 4.5 foundation, X1.1 uses what they call a “hybrid reinforcement learning framework”, basically combining different AI training techniques to make the model smarter.

The technical details might sound dry, but the results Baidu is claiming are pretty impressive. They say X1.1 is significantly better than its predecessor:

• 34.8% more accurate with facts
• 12.5% better at following instructions
• 9.6% improvement in handling complex tasks

The Big Claims

Here’s where things get interesting (and a bit bold). Baidu isn’t just comparing their new model to their own previous work; they’re taking shots at the competition. They claim Ernie X1.1 outperforms DeepSeek’s R1-0528 model across multiple tests.

But wait, there’s more. They’re also saying their model can go toe-to-toe with some of the biggest names in AI, specifically mentioning performance comparable to GPT-5 and Google‘s Gemini 2.5 Pro in certain areas.

Where Can You Try It?

If you’re curious to test it out, Baidu has made X1.1 available through their usual channels – the Ernie Bot website and their Wenxiaoyan app. For businesses and developers, they’ve also deployed it on their Qianfan AI Cloud platform.

The Bigger Picture

This launch isn’t happening in isolation. Baidu also updated their PaddlePaddle deep learning framework to version 3.2, making it better for training large AI models and working with different types of hardware. They’ve been busy updating their development tools too.

The numbers they’re throwing around are pretty staggering – apparently, their AI ecosystem now supports over 23 million developers and serves 760,000 companies. That’s a massive user base, especially concentrated in China.

What This Really Means

Let’s be honest – every AI company is making big claims about their latest models these days. But Baidu’s push with X1.1 represents something important: the global AI race is heating up, and Chinese companies aren’t content to just follow along.

Whether Ernie X1.1 actually lives up to these bold comparisons with GPT-5 and Gemini remains to be seen. Independent testing and real-world use will be the real judge. But one thing’s for sure, Baidu is clearly positioning itself as a serious player on the global AI stage, not just in China.

The competition between AI giants is only getting more intense, and that’s probably good news for everyone who uses these tools.

In related news, Baidu has recently introduced GenFlow 2.0, its latest AI agent framework designed to deliver real-world task execution.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

(Via)

The post Baidu Unveils Ernie X1.1 Deep Thinking Model, Claims It Outperforms DeepSeek R1 appeared first on Gizmochina.

Embodied intelligence refers to the ability of robots to combine physical actions with advanced reasoning, perception, and decision-making. Unlike traditional machines that follow fixed commands, embodied robots can interpret complex instructions, adapt to new situations, and perform tasks with human-like flexibility. This appears set to become a reality in the not-too-distant future, as an increasing number of players and nations are entering this competitive arena. Currently, the race for embodied intelligence is intensifying.

The Embodied Intelligence Race

Tesla is placing huge weight on its humanoid robot project, Optimus. The company expects it to become a major part of its future, representing up to 80% of Tesla’s overall value. Elon Musk has set an ambitious target: producing one million humanoid robots per year by 2030. Optimus is designed primarily for factory use but is also being developed for consumer applications, ranging from household assistance to personal support.

Boston Dynamics, long known for its groundbreaking robotic designs, has turned its Atlas model into a fully electric humanoid platform. Atlas is capable of performing complex multi-task demonstrations, such as handling tools or navigating obstacle courses. With Toyota’s large Behavior Model supporting its development, Atlas is being positioned as an advanced industrial solution. However, like Tesla’s Optimus, Atlas still faces the challenge of moving beyond staged demonstrations toward everyday usability.

NVIDIA is entering the robotics race with a software-first approach. Its Isaac platform and GR00T foundation models provide robotics developers with tools for simulation, reasoning, and adaptability. NVIDIA describes GR00T as “the next wave of AI,” aiming to make robots more capable of decision-making in real-world settings. This positions NVIDIA as a key enabler of robotics rather than a hardware maker, offering the “brains” for embodied intelligence.

The Challenge: Demos vs. Dependability

Despite impressive demonstrations of robots running, jumping, or manipulating tools, most humanoid robots struggle with real-life tasks. Simple activities like folding laundry, serving food, or caregiving remain much harder than advertised. The gap between marketing demos and practical dependability is still wide, slowing adoption in homes and industries.

China’s Disruptive Entry: X Square Robot and Wall-OSS

A new player from China, X Square Robot, is taking a bold step with its open-source model, Wall-OSS. This is China’s first foundational model for embodied intelligence and is designed to make robots adaptable in unpredictable real-world environments. Unlike proprietary approaches from Tesla or Boston Dynamics, Wall-OSS will be freely available on platforms like GitHub and Hugging Face.

How Wall-OSS Works

Wall-OSS uses a Shared Attention Mechanism to focus only on relevant cues, improving reaction time and reducing errors. It’s also paired with a Task-Routed Feed-Forward Network (FFN). In most legacy systems, all sensory streams, such as camera vision, spoken instructions, or motor commands, are funneled through a single processing layer, forcing the model to juggle unrelated inputs in one space. The result was bottlenecks, slow adaptation, and a tendency to misprioritize commands.

Wall-OSS addresses this differently. Shared attention allows the model to selectively focus on the most relevant cues in a scene, while the task-routed FFN processes different types of input, vision, language, and motor actions through specialized pathways. For example, visual data is processed along one optimised pathway for object recognition and spatial mapping; linguistic commands are parsed and processed through a separate pathway. Concurrently, motion processing operates independently, accounting for physical constraints and real-time feedback.

This enables robots to understand and act in context, such as following the instruction “Pick up the apple on the table and place it in the bowl” instead of treating “see apple” and “pick” as separate tasks. Routing multimodal inputs effectively mirrors human cognition, where vision/sign, hearing, and motor planning seamlessly integrate. For robots, this translates to faster response times, fewer errors, and better performance in unfamiliar environments.

Another key feature is Chain-of-Thought (CoT) reasoning integrated into its architecture, where the robot plans out multi-step actions before execution. For example, a command like “clean the table” is often treated as a one-off trigger by robots in traditional models. They may forget to wipe it down after picking up tableware. While CoT reasoning gives the ability to simulate multiple steps before moving. It generates an internal plan instead of reacting in isolation. Responding to the command ‘clean the table’ involves recognizing clutter, sorting items, removing dishes, and wiping the surface, all done logically rather than through trial and error.

This, of course, extends beyond simple housekeeping. In industrial or service settings, multi-step planning allows robots to adapt without explicit programming. A warehouse bot could deduce how to stack differently-sized packages without knocking them over, while a healthcare assistant could prepare instruments by following procedural sequences.

Training for Real-World Adaptability

Unlike older systems trained on narrow datasets, Wall-OSS has been trained on billions of Vision-Language-Action samples. These come from real-world robotic logs, generative videos, and synthetic environments with different lighting, textures, and clutter. This broad training gives model resilience beyond controlled labs, and also means that WALL-OSS-powered robots are less likely to falter when faced with unusual household layouts, different object shapes, or sudden changes in context.

The Quanta X2 Robot

To showcase Wall-OSS, X Square Robot has developed the Quanta X2. This robot uses a wheeled base, a 7-degree-of-freedom arm, and a dexterous hand with lifelike gestures. It can handle up to 62 degrees of freedom for natural movements and features rotating clamps for 360° cleaning. The Quanta X2 is designed for use in service industries, households, and industrial settings.

Open-Source vs. Proprietary Futures

While other robotics companies maintain proprietary stacks, X Square Robot is set to release Wall-OSS openly on GitHub and Hugging Face. And if adoption spreads, Wall-OSS may position open frameworks as a credible counterpoint to closed systems.

With roughly US $100 million in funding, X Square Robot is betting that open-source collaboration can solve what proprietary efforts have not: dependable performance in unpredictable environments. For startups, Wall-OSS provides a ready-made brain; for hardware makers, it offers a generalizable intelligence layer. The race is no longer about choreographed performances but about bridging the gap between shallow demos and real-world needs. 

The post Embodied Intelligence Race: How China’s Open-Source Robotics Is Challenging Global Leaders appeared first on Gizmochina.

Chinese startup DeepSeek is preparing to launch a next-generation AI agent by the end of 2025, according to a Bloomberg report. The Hangzhou-based company is developing a system that can perform complex, multi-step tasks with minimal input from users. The model will also learn from its past actions to improve performance over time.

DeepSeek gained global attention earlier this year with the release of its R1 model. The system demonstrated advanced reasoning capabilities and was built at a reported cost of just $6 million, far less than rivals like OpenAI’s ChatGPT or Google’s Gemini. R1 was also open source, allowing developers to freely access its code. The release disrupted the “bigger is better” approach in Silicon Valley, where massive budgets and vast datasets have dominated AI development.

Since R1, DeepSeek has moved cautiously. Founder Liang Wenfeng delayed the R2 model, a move attributed to technical refinements and his ongoing work at High-Flyer Asset Management. This deliberate pace differs from competitors like Alibaba and Tencent, which have accelerated their own AI rollouts.

Last month, DeepSeek rolled out its V3.1 update, expanding the context window to 128,000 tokens and increasing the parameter count to 685 billion. The company also now applies mandatory AI-generated content labels on all outputs, a policy that cannot be disabled and sets DeepSeek apart from its peers.

The upcoming R2 model will target the emerging market for AI agents. Unlike chatbots that deliver text-based responses, agents are designed to handle tasks such as travel planning, software debugging, and business workflows with little human supervision. Analysts view this technology as the next major step in AI, capable of boosting productivity and reshaping digital services.

Global competitors, including OpenAI, Microsoft, and Anthropic, have already introduced agent-focused features. DeepSeek aims to match and potentially surpass those efforts, continuing its push to challenge U.S. dominance in AI.

Details about pricing, availability, and technical specifications remain limited. However, industry watchers expect the launch to be closely monitored in both Silicon Valley and Washington, where DeepSeek’s rapid progress has already raised concerns about China’s growing influence in artificial intelligence.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

(Source)

The post DeepSeek is working on AI agent to take on OpenAI by end of 2025 appeared first on Gizmochina.

DeepSeek, the AI company, has just rolled out new rules in China requiring all AI-generated content on their platform to be clearly marked as artificial. This change comes as part of broader efforts to make AI content more transparent and prevent people from misusing it.

The company sent out an official notice explaining that anything created by their AI systems must now include clear signs showing it wasn’t made by humans. These markers come in two forms: obvious ones that people can easily see, and hidden technical ones built into the content itself.

The visible markers might be text saying “AI-generated,” audio announcements, or graphics that appear somewhere in the content. The hidden markers are tucked away in the technical data and include details like what type of content it is, which company made it, and a unique ID number.

Users aren’t allowed to mess with these labels in any way. They can’t remove them, change them, fake them, or try to hide them. DeepSeek has also banned any tools or services that would help people tamper with the labels. If someone breaks these rules, they could face legal consequences.

Along with this announcement, DeepSeek also released a detailed technical guide explaining how their AI actually works. The document covers things like how they train their models, what data they use, and the process behind creating content. They say this is meant to help people understand the technology better, use it more responsibly, and give everyone the right to know where content comes from.

These new requirements aren’t just DeepSeek’s idea. They are following the new Chinese government regulations. The rules were put in place by several government agencies and require all AI-generated content to have labels that can be tracked and cannot be tampered with. The companies providing these AI services are responsible for making sure everyone follows the rules.

In related news, DeepSeek has recently released V3.1 with a 128K token context and 685B parameters, unifying reasoning and general tasks while the R2 model remains delayed. Meanwhile, Nvidia’s GB300 GPUs reportedly achieve 6x the performance of H100s on DeepSeek R1, offering major efficiency and scalability gains for enterprise AI.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

(Via)

The post DeepSeek now slaps AI labels on all content and you can’t remove them appeared first on Gizmochina.

DingTalk, Alibaba’s enterprise communication and collaboration platform, unveiled its first AI hardware device, the DingTalk A1, at its 10th anniversary event. The company released an initial batch of 1,000 units at the venue, which sold out immediately.

DingTalk A1 Specifications

The DingTalk A1 is a slim AI-powered recording card that works with the newly launched AI Note system. It records, transcribes, translates, summarizes, and analyzes conversations. The device connects directly with the DingTalk app, making it a pocket-sized AI assistant for professionals.

The hardware targets roles such as sales teams, HR, lawyers, teachers, doctors, and customer service staff. It automatically turns spoken conversations into digital notes, summaries, and structured insights that can be shared within DingTalk. The device includes over 30 scenario-based templates covering meetings, interviews, consultations, and classroom use.

The A1 measures 3.8mm thick and weighs 40.8g. It attaches magnetically to the back of a smartphone for call recording and quick memos. The device uses a BES2800 6nm AI audio chip from Hengxuan Technology and features a 660mAh battery. It supports up to 45 hours of continuous recording or 60 days on standby. Local storage capacity is 64GB.

The flagship version has an aluminum alloy body, a small front display, and two control buttons. It comes with six microphones, including one bone conduction microphone, that capture voices within 8 meters. The system can reduce more than 500 types of background noise. Recordings are encrypted across device, app, and cloud, with enterprise options for private deployment and centralized data management.

The device also enables real-time translation, with text shown on the DingTalk app or projected on a larger screen. In interviews, HR teams can receive automatic summaries of candidates’ profiles and communication abilities. In meetings, it can generate mind maps and organize customer information.

Pricing and Availability

DingTalk launched two versions of the device. The flagship edition costs 799 yuan ($112) and includes unlimited speech-to-text conversion, unlimited cloud storage, support for 120 languages, and unlimited AI-generated analyses. The youth edition costs 499 yuan ($70) and has a leather-like finish without a front display. It includes 1,000 minutes of transcription and 10GB of cloud storage per month, with support for three languages.

The first batch of flagship models sold out on Tmall. DingTalk will begin open sales on September 15 through its official flagship store.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

(Via)

The post This Credit Card-Sized AI Recorder Could Replace Your Meeting Notes Forever appeared first on Gizmochina.

Tesla has introduced a new AI-powered voice assistant in its electric vehicles in China, developed in partnership with DeepSeek and ByteDance. The feature is called “Hey Tesla” and is designed to give drivers natural voice control for navigation, media, and cabin functions.

Tesla confirmed that ByteDance’s Doubao large language model will manage voice commands such as adjusting the air conditioning, setting routes, and controlling entertainment. DeepSeek’s chatbot will handle conversational tasks, including news, weather updates, and casual interaction. Both models operate on ByteDance’s Volcano Engine cloud platform, which processes requests in real time through an encrypted API.

The company launched the system with the new six-seat Model Y L SUV, which went on sale in August. Unlike older Tesla models in China that require pressing a steering wheel button, the Model Y L supports a wake phrase. Drivers can say “Hey Tesla” or set a custom word to activate the assistant hands-free.

Tesla’s choice of partners reflects its push to localize features for the Chinese market. Domestic rivals such as BYD, Nio, and Xpeng already use advanced AI voice assistants, giving them an advantage in customer experience. By adopting DeepSeek and Doubao, Tesla aims to close that gap and compete more directly with homegrown brands.

In the United States, Tesla continues to use Grok, the AI model built by Elon Musk’s startup xAI. The split shows Tesla’s regional strategy, with local AI providers powering its cars in China to meet regulatory and consumer expectations. BMW has taken a similar path, working with Alibaba’s QWen LLM for its Chinese lineup.

Tesla has not confirmed when the new voice assistant will expand to other models in China. The company’s updated terms of service mention the integration, but some owners said they had not received the feature in recent over-the-air updates. Despite the delay, Tesla is signaling that in-car intelligence is now central to competing in the world’s largest EV market.

In related news, we recently covered a comprehensive guide on AI chatbot subscriptions in 2025, including all the key plans, pricing, and features you need to know.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

(Via)

The post Tesla partners with DeepSeek and ByteDance to launch AI voice assistant in China appeared first on Gizmochina.

Chinese AI company DeepSeek has released version 3.1 of its flagship large language model, expanding the context window to 128,000 tokens and increasing the parameter count to 685 billion. The update was announced quietly through the company’s WeChat user group on August 19, without any posts on its official social media channels.

What’s New?

The most significant change in DeepSeek V3.1 is the increased context length, which now allows the model to handle inputs equivalent to a 300 to 400-page book. This improvement enables better performance in long-form content generation, technical document analysis, and extended multi-turn conversations. The official group confirmed the expanded context was already supported internally in the previous V3 version, but has now been formally enabled across all interfaces.

Mixture-of-Experts and Benchmark Scores

DeepSeek V3.1 continues to use a Mixture-of-Experts (MoE) architecture, with only 37 billion parameters activated per token. The model supports multiple precision formats, including BF16, FP8, and F32, providing more flexibility for different deployment environments. Developers can access the model via API or download it from Hugging Face under the MIT open-source license.

The upgraded model performed well in early third-party benchmarks. It scored 71.6% on the Aider coding test, placing it above Claude Opus 4 and making it one of the strongest open-source coding models currently available. DeepSeek V3.1 also showed improved performance in math and logic tasks. However, some users noted no clear gains in reasoning compared to the earlier R1-0528 model.

Shift in Strategy

DeepSeek has removed all references to the R1 model from its chatbot interface, signaling a shift toward a single hybrid model architecture. The company appears to have integrated its reasoning capabilities into V3.1, instead of maintaining a separate reasoning model.

The training cost for V3.1 has not been disclosed. However, according to previous reports, the original V3 model was trained on 2.788 million GPU hours using Nvidia H800 chips, at an estimated cost of $5.6 million. That model formed the base for the current version, which likely shares similar infrastructure with additional refinements.

Confusion Around the Delayed R2 Model

There had been widespread anticipation that DeepSeek’s next major release would be the long-awaited R2 model, designed to advance reasoning capabilities. Instead, V3.1 appeared as the company’s next step. According to a recent Financial Times report, the R2 model’s release has been delayed due to persistent technical issues involving Huawei’s Ascend AI chips.

DeepSeek was reportedly urged to use Ascend hardware to reduce reliance on Nvidia, aligning with China’s national strategy for AI self-sufficiency. Despite support from Huawei engineers, training on Ascend failed due to compatibility and performance issues. The company then switched to using Nvidia GPUs for training while retaining Ascend for inference. This hybrid setup introduced further complications and delays. In addition, the extended time spent on data labeling slowed down development. DeepSeek founder Liang Wenfeng reportedly expressed frustration with the slow progress.

Meanwhile, competitors such as Alibaba’s Qwen3 have moved ahead by deploying similar algorithms with more efficient execution. The episode has underscored the limitations of China’s domestic chip infrastructure and the challenges faced by startups attempting to meet political and technical demands simultaneously.

DeepSeek has not ruled out the launch of R2. However, whenever the model does arrive, its performance will face intense scrutiny. Until then, V3.1 stands as the company’s current flagship, serving both reasoning and non-reasoning workloads in a unified framework.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

The post DeepSeek Releases V3.1 Model: What’s New? appeared first on Gizmochina.

Baidu is making some bold claims about its latest AI creation. At Baidu AI Day Open Day, the Chinese search and cloud computing giant unveiled GenFlow 2.0, which it’s billing as “the world’s first all-platform general-purpose intelligent agent.” The system promises to solve one of AI’s most persistent problems: actually completing complex, multi-step tasks without making users wait around or deliver disappointing results.

100 agents, three minutes, bold promises

The pitch is compelling, if ambitious. GenFlow 2.0 can supposedly orchestrate more than 100 specialized AI agents working in parallel, completing over five complex tasks simultaneously within a three-minute window. Baidu claims the system is 10 times faster than competing platforms, delivering what it calls “minute-level output” with processes that humans can interrupt and modify as needed.

Unlike many AI agent platforms that require invitation codes or put users in queues, GenFlow 2.0 is immediately available to anyone through Baidu Wenku’s web and mobile platforms. This accessibility could be crucial for adoption, especially as users grow frustrated with AI tools that promise the world but gate access behind waitlists and beta programs.

The technical complexity behind the curtain

The technical architecture behind GenFlow 2.0 sounds appropriately complex for its ambitious goals. Baidu has built a proprietary Multi-Agent engine paired with an optimized context management system designed to create what the company calls an “information ecosystem for the model.” The platform uses dynamic hybrid inference and multi-mode intelligent scheduling to handle complex workflows, while incorporating both public knowledge and private user data to enhance results.

Perhaps more importantly for real-world usage, GenFlow 2.0 includes systems for multimodal data understanding, both short and long-term memory, and multi-level risk controls. The platform also supports the Model Control Protocol (MCP), which should allow third-party services to integrate more easily with Baidu’s ecosystem.

Honor partnership brings agents to your phone

The most interesting partnership announcement involves Honor. Honor has become one of the first hardware partners to join Baidu’s MCP ecosystem, natively integrating GenFlow 2.0 into its YOYO smart assistant. This integration could give Honor’s MagicOS users system-level AI agent capabilities, allowing them to access their Baidu Netdisk files, search through documents, generate presentations, and perform other complex tasks through voice commands or touch interfaces.

The scale of Baidu’s content ecosystem provides a significant advantage here. Baidu Wenku hosts over 1.4 billion pieces of professional content and serves 97 million monthly active AI users, while Baidu Netdisk claims more than 1 billion total users, including 200 million monthly actives and 80 million AI users. This massive content repository could give GenFlow 2.0’s agents a substantial knowledge base to work with.

The reality check on AI agent promises

This launch builds on Baidu’s previous “Cangzhou OS,” which the company described as the world’s first content-centric operating system and formed the foundation for the original GenFlow platform. That system already supported multimodal input and output, multi-agent coordination, and MCP compatibility.

The real test will be whether GenFlow 2.0 can deliver on its promises in daily use. AI agents have consistently struggled with reliability, accuracy, and the ability to handle complex, ambiguous instructions. If Baidu has cracked the code on fast, reliable multi-agent coordination, it could represent a significant leap forward for practical AI applications. But given the industry’s track record of overpromising and underdelivering on AI agents, users should probably manage their expectations while remaining cautiously optimistic.

In related news, China is turning to AI to address its legal service gaps, with a new model trained on 200 million cases aiming to assist underserved regions.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

(Via)

The post Baidu’s GenFlow 2.0 wants to be the AI agent that actually gets things done appeared first on Gizmochina.