AI News

AI changes every day,
catch up on it all on one page

The latest updates from AI companies, research institutions, and media around the world, automatically selected for you and updated around the clock.

Updated 11/16 19:04 (UTC+8) · refreshed hourly

Today11/16 · Monday4 items

Tencent Hunyuan: ResearchModelsAI Score73

Tencent Hunyuan Releases Hy4 flash: A Cost-Effective Everyday Model with 224B Total Parameters and a 1M Context Window

Tencent Hunyuan introduces Hy4 flash, with 224B total parameters (including 64B Engram parameters) and 12B activated parameters. It supports multimodal input and a 1M context window. API pricing is much lower than Hy4 (0.5 yuan for input and 2 yuan for output per million tokens).

Why it's recommended:The original article provides the architecture, pricing table, cost comparisons, and evaluation data across multiple business scenarios, helping readers assess cost-effective options for everyday tasks.

Also reported by 3 other sources
Arena.ai@arenaModelsAI Score82

Qwen4-Max tops the Chinese model rankings in Agent Arena, with Alibaba rising to No. 14 overall

The Arena rankings show Qwen4-Max at No. 14 overall in Agent Arena with a +7.5% net improvement, making it the top Chinese model. Alibaba consequently becomes the highest-ranked Chinese lab and No. 5 overall, behind only Anthropic, Google, OpenAI, and Meta.

Why it's recommended:Based on 14.7k real Agent sessions, the rankings provide placements and a breakdown of four signals, allowing readers to compare Qwen4-Max's actual performance with other Chinese models.

Also reported by 1 other source
Qwen@Alibaba_QwenModelsAI Score73

Alibaba Qwen Releases Qwen4 Max, Scoring 9 Points Higher Than Qwen3.8 Max on the Intelligence Index at Lower Cost

Alibaba Qwen announced the release of Qwen4 Max. According to Artificial Analysis, it scores 43 on the Intelligence Index, 9 points higher than Qwen3.8 Max (xhigh) and 3 points behind Kimi K3.1 Pro. Its cost per task is $3.27, about one-third that of Qwen3.8 Max, while output token usage is reduced by 39%. It has a 1M-token context window and is priced at $1.50/$4.50 per 1M input/output tokens. It is available through Alibaba's first-party API.

Why it's recommended:The original article uses Artificial Analysis data to compare Qwen4 Max's scores and cost changes, helping readers assess its value relative to its predecessor and competitors.

Also reported by 1 other source
Qwen:Blog RetrievalModelsAI Score80

Qwen releases Qwen4-Max: a 2.4T-parameter backbone, a million-token context window, and long-horizon Agent capabilities

The Qwen team released Qwen4-Max, based on the new Qwen4 architecture, with 2.4T backbone parameters, 71B parameters activated per token, and 328B N-gram Embedding parameters. It has native vision capabilities and a 1 million-token context window.

Why it's recommended:The official release also provides architectural details, cost comparisons, and real-world industry cases, allowing readers to assess the actual changes in Qwen4-Max compared with the previous generation.

Reported by 6 additional sources

11/13 · Friday8 items

OpenMOSS / Fudan NLP / Shanghai Innovation Institute / MOSIModelsAI Score60

Understanding MOSS-Live’s full-duplex real-time interaction evaluation: A comparison with GPT-Live-1 and Doubao

The OpenMOSS team published an analysis of MOSS-Live’s full-duplex voice interaction evaluation. According to a chart dated November 13, 2026, on the official MOSS website, MOSS-Live ranked first among the evaluated models on all six full-duplex metrics. Its false response rate to irrelevant speech was 2.00%, compared with 68% for GPT-Live-1 and 75% for Doubao, while its interruption rate during pauses was 1.44%.

Why it's recommended:The article compiles full-duplex evaluation comparisons between MOSS-Live and three voice models using charts from the official website, while also highlighting remaining gaps in multi-turn tasks and comprehension metrics.

Official account: Tencent HunyuanModelsAI Score66

Tencent Hunyuan releases Hy Realtime voice model with tool calling and singing capabilities

Tencent Hunyuan has released its new real-time voice model, Hy Realtime. In addition to core capabilities such as speech comprehension, instruction following, and humanlike expression, it adds agentic tool calling and singing. It is already available in Yuanbao voice calls and Tencent Customer Service, and will soon launch on Tencent Cloud with API access.

Why it's recommended:The official announcement provides full training details covering cross-modal alignment, the data system, and reinforcement learning, along with products already using the model and links to try it, helping readers understand its practical capabilities.

Also reported by 1 other source
Meituan LongCat: New Hugging Face modelModelsAI Score61

Meituan LongCat open-sources LongCat-Video-Avatar 1.6, an audio-driven digital human video generation model

Meituan’s LongCat team has released LongCat-Video-Avatar 1.6, an open-source model for audio-driven human video generation. It improves the naturalness of mouth movements, lip synchronization, coordination of facial and body movements, and visual consistency in long videos. It reduces error accumulation during autoregressive continuation, can generate videos longer than 10 minutes, and adds end-to-end inference optimizations.

Why it's recommended:The original article details the improvements in version 1.6, rankings from a human evaluation with 33,051 votes, and complete deployment parameters, allowing readers to assess the extent of the upgrade and inference configurations.

Hacker News: Popular AI postsModelsAI Score76

Edge0 open-sources a 145B on-device model that can run locally on a laptop

Edge0 announced the open-source release of the Edge0-145B and Edge0-120B models, along with a unified inference framework and deployment tools. It describes Edge0-145B as the first open-source model with 145B parameters that can run locally on a personal computer.

Why it's recommended:The original article provides specific benchmark scores and measured memory usage and speed, alongside open-source weights and an inference framework, allowing readers to assess the practical usability of large on-device models.

Google Gemini@GeminiAppModelsAI Score71

Gemini 4 Hydrogen begins rolling out to all users worldwide today

Google Gemini announced that Gemini 4 Hydrogen begins rolling out to all users worldwide today across desktop, mobile, and web. Free users do not need to select the model separately: automatic model selection can route prompts to it when appropriate. Details are available at https://goo.gle/4AxH40q.

Why it's recommended:The official announcement explains how free users can access Gemini 4 Hydrogen, helping readers understand its availability and automatic model selection mechanism.

Google@GoogleModelsAI Score76

Gemini 4 Hydrogen begins rolling out to everyone across the Gemini App, Search, and Gemini API

Google announced that Gemini 4 Hydrogen begins rolling out today. All users can access it in the Gemini App, Google Search, and Google AI Studio. AI Pro and Ultra subscribers can also use it in Antigravity, Gemini in Chrome, and Gemini in Google Sheets. Developers and enterprises can access it through Google AI Studio, AndroidStudio, and all Gemini API tiers.

Why it's recommended:The original article details the full availability of Gemini 4 Hydrogen across consumer, subscription, and developer channels, making it easy for readers to identify their access options.

Artificial Analysis@ArtificialAnlysModelsAI Score78

Artificial Analysis evaluates Gemini 4 Hydrogen: Intelligence Index score of 54, close to Sonnet 5.5 at approximately 32% of its cost

Artificial Analysis evaluated Google’s newly released Gemini 4 Hydrogen. Its new xhigh reasoning level scored 54 on the Intelligence Index, 1 point below Claude Sonnet 5.5 (max) and GPT-6.1 Sol (max). Its cost per task is approximately 32% of Sonnet’s, but approximately 2.5 times that of GPT-6.1 Sol.

Why it's recommended:The evaluator provides intelligence, hallucination rate, terminal task performance, and cost data for Gemini 4 Hydrogen at each reasoning level, helping readers compare prices and select a model for their use cases.

Google DeepMind:BlogModelsAI Score82

Google DeepMind releases Gemini 4 Hydrogen

Google DeepMind has released Gemini 4 Hydrogen, positioned for fast execution of everyday coding tasks and agent workflows, with availability across all tiers through the Gemini API, AI Studio, the Gemini app, Google Search, and other channels.

Why it's recommended:The official announcement provides multiple benchmark comparisons with similar models and pricing details, helping readers assess its role in workflows involving multiple models.

Reported by 14 additional sources

11/12 · Thursday1 item

Runway:NewsModelsAI Score67

Runway releases open weights for 27B world-action model Praxis-1 under the Apache 2.0 license

Runway announced that the weights for its world-action model Praxis-1 are available to everyone under Apache 2.0. The model is a 27B-parameter MoE diffusion Transformer that jointly predicts video and actions. It tops the NVIDIA RoboLab-120 leaderboard with a success rate of 46.2%, ahead of FLUX 3 Action (42.9%), Cosmos3 Nano (36.8%), and π0.5 (28.0%).

Why it's recommended:The original article provides specific RoboLab-120 leaderboard results, details of what has been open-sourced, and adaptation methods, helping readers assess the feasibility of deployment on their own hardware.

Also reported by 2 other sources

11/11 · Wednesday1 item

Baidu ERNIE official blogModelsAI Score73

Baidu releases ERNIE 5.5 Flash Preview, with weights to be open-sourced by November 25

Baidu has released ERNIE 5.5 Flash Preview, the first post-trained version based on its earlier Base model. It has 280B total parameters, 14B active parameters, and support for a 1M context window. The model API is now available on the Qianfan platform, and its weights will be open-sourced by November 25. The official announcement showcases results in agent scenarios including game development, 3D modeling, knowledge work, and video creation, and describes this release as an intermediate post-training version, with further optimization planned for the final version.

Why it's recommended:The official specifications and benchmark comparisons include 280B total parameters, 14B active parameters, a 1M context window, and more, providing a basis for evaluating the model's positioning in coding and agent tasks.

Also reported by 1 other source

11/10 · Tuesday9 items

Google DeepMind:BlogModelsAI Score85

Google DeepMind releases Gemini 4 Carbon, a frontier model

Google DeepMind has officially made Gemini 4 Carbon available to developers, enterprises, scientists, and general users. Positioned as its most capable frontier model, it focuses on cybersecurity and long-horizon agent workflows in complex knowledge work, software engineering, and scientific reasoning.

Why it's recommended:The official release page provides benchmark rankings, pricing, and safety data, enabling readers to assess the respective roles and applicable workflows of Carbon and Argon.

Reported by 12 other sources
MiniMax:BlogModelsAI Score66

MiniMax releases the open-source image generation model H3-Image, generating 4K images in under 10 seconds

MiniMax has released H3-Image, an open-weight image generation model based on H3. It supports text-to-image generation, reference-image generation, and image editing, can generate 4096 × 4096 4K images, and reaches speeds up to 5 times those of mainstream solutions.

Why it's recommended:The original article provides specific metrics for speed, image quality, and prompt adherence, along with access to the open weights, enabling developers to assess its usability in their own workflows.

Reported by 4 other sources
The Decoder:AI NewsModelsAI Score81

Anthropic Fable 5.2 leads real freelance tasks on the Remote Labor Index with 41.3%

The Decoder reports that Anthropic's new model Fable 5.2 ranked first with 41.3% on the Remote Labor Index benchmark from Scale Labs and the Center for AI Safety. The test measures whether AI agents can complete real freelance projects to a quality level acceptable to paying clients.

Why it's recommended:The original article compiles the complete Fable 5.2 leaderboard and task details for the Remote Labor Index, allowing readers to compare the actual performance gaps among models on real client-work tasks.

WeChat Official Account: Digital Life KazikeModelsAI Score81

Hands-on testing Claude Fable 5.2: one-shot video and 3D web pages, with standout aesthetics and detail

Anthropic officially released Claude Fable 5.2 in the early morning. The author immediately tested 10 tasks, including video motion effects, Three.js 3D scenes, and a 3D web page for Apple's Jing'an store. All were generated successfully in one shot, and no factual errors were found in the generated videos.

Why it's recommended:Using multiple one-shot testing examples, the author demonstrates Fable 5.2's performance in video motion effects and 3D front-end development, and provides recommendations on quota usage and everyday model combinations.

Zyphra ResearchModelsAI Score65

Zyphra releases the open-source TTS model ZONOS3-Mini, which can run in real time in a mobile browser

Zyphra has released ZONOS3-Mini, an Apache 2.0 open-source zero-shot voice-cloning TTS model. It uses a sparse MoE architecture with 800 million total parameters and 100 million active parameters. Its Time to First Audio is under 570 ms in the Pixel 9 browser and under 150 ms on an M3 MacBook Air.

Why it's recommended:The original article provides specific data on parameter size, on-device latency, and evaluation rankings, allowing readers to assess this open-source TTS model's practical positioning in voice cloning and controllability.

Also reported by 1 other source
Artificial Analysis@ArtificialAnlysModelsAI Score81

Artificial Analysis evaluates Claude Fable 5.2: tops the Intelligence Index v5 with a score of 65

Artificial Analysis evaluated Anthropic's Claude Fable 5.2, which topped the Artificial Analysis Intelligence Index v5 with a score of 65, ahead of Gemini 4 Carbon (60) and Claude Opus 5.5 (59).

Why it's recommended:The evaluator provided Claude Fable 5.2's score on the Intelligence Index v5, benchmark comparisons, and cost data, enabling readers to assess its capabilities and cost-effectiveness.

Load more

Finished reading the news? Let’s talk about your business.

Leave your contact details, and I’ll get in touch with you personally.

WECHAT

Add me on WeChat

Press and hold or select the WeChat ID below to copy it.