Tencent Hunyuan discussed the Hy4 model family with developers during a Reddit AMA, according to 智东西. The team said the next Hy model will expand in scale, identified Hy4-mini, Flash-Lite and models for 24GB GPUs as planned work, and disclosed architecture choices including MLA, DSA, iHC, LatentMoE and Engram. It also said Tencent will continue releasing open-weight models while restricting some o…
Tencent Hunyuan announced that the weights for its Hy4 large language model are now available on GitHub, Hugging Face, ModelScope, and GitCode for download and deployment. IT之家 reports that this completes the open-sourcing of the Hunyuan LLM series, including Hy3 preview, Hy3, Hy4 preview, Hy4, and Hy4 flash. Tencent says Hy4 uses a 770B-parameter MoE architecture with 49B active parameters and a …
Google DeepMind began rolling out Genie 4 to eligible Google AI Ultra subscribers through Project Genie. The model supports real-time natural-language environment editing, native adaptive audio, and sharing through a community gallery, alongside a Project Genie agent and updates to Genie 3. Google AI Pro members can join the waitlist.
Why it matters: Genie 4 turns Google’s experimental world-mode…
Tencent released and open-sourced the Hy4 Flash family, comprising Hy4 Flash, Hy4 Flash-FP8 and Hy4 Flash Base, with downloadable weights for local deployment. The models natively accept text and images, support context windows of up to 1M tokens and activate 12B parameters per token; Tencent’s materials describe total size as either 224B or 220B. In PingWest’s WorkBuddy comparison, Hy4 Flash used…
Microsoft released Microsoft-Decision-1.1 on November 17, 2026, making it available through Microsoft Foundry, OpenRouter and Vercel. The model is designed for routing, classification, content verification and workflow control, selecting permitted actions within developer constraints and returning calibrated confidence scores. Artificial Analysis reported a score of 68 on its Decision Index v2.1, …
Tencent Hunyuan released Hy4 flash, a multimodal language model with 224B total parameters, 12B active parameters per token, and a 1M-token context window. According to 36氪, API pricing is RMB 0.5 per million input tokens and RMB 2 per million output tokens, with cached input priced at RMB 0.02; IT之家 reports that the model is already available through Tencent Cloud TokenHub, WorkBuddy, Yuanbao, We…
Alibaba released Qwen4-Max, a foundation model it says has 2.4 trillion total parameters, native vision capabilities, and a million-token context window. The model is available through the 千问AI platform via API, Alibaba Token Plan, and several Qwen applications, while Alibaba plans upcoming open-source releases of smaller Qwen4 models.
Why it matters: Qwen4-Max expands Alibaba’s Qwen model lineup…
Tencent Hunyuan released Hy Realtime, a real-time voice model with conversational understanding, instruction following, expressive speech, tool calling and singing. Tencent says it can reconstruct intent from incomplete or corrected speech and is already available in Yuanbao voice calls and Tencent customer service, with a Tencent Cloud API planned. Users can try the current model through the late…
Google has brought Gemini 4 Hydrogen to Google Search globally, enabling improved intent understanding, simultaneous queries, and structured results such as tables, images, and links. The model is also available through the Gemini App, including for free users through automatic model selection, as well as Google AI Studio and the Gemini API, according to Google and subsequent reporting.
Why it ma…
Google announced the Gemini Decisions API, a developer endpoint powered by Gemini Decision 1 for closed-set decisions across text, images, audio, video, and PDFs. It can score up to 1,000 yes/no, multiple-choice, and rubric questions in one request, returning probabilities and confidence scores; developers can access it through Google AI Studio. Google says the model delivers 150–250 ms end-to-end…
Google DeepMind unveiled AlphaGenome 2, a model designed to predict how genetic variants affect specific cell types and states. The company says it was trained on more than 100 million individual cell profiles spanning 68 tissues and identifies nearly twice as many variants that can alter gene expression as its predecessor, though these performance claims are company-reported. Google is making the…
Alibaba previewed Qwen4-Max through Alibaba Token Plan on November 11, 2026, while the full Qwen4 release date remained undisclosed. Qwen said the new generation advances knowledge understanding, world perception, coding and agentic capabilities. IT之家 and AIbase reported that Qwen Office, Qoder, Qwen App and QwenStudio had also integrated the new model. The reports did not disclose model parameter…
Google DeepMind announced on November 10, 2026 that Gemini 4 Carbon is available to developers, enterprises, scientists and consumers through paid API access and several Google products. The company positions Carbon as a frontier model for long-horizon agentic workflows, complex coding, cybersecurity and scientific reasoning, claiming first-place rankings on CursorBench 4.0, Perplexity’s WANDR and…
TestingCatalog reports that Anthropic launched Claude Fable 5.2 and Claude Mythos 5.2, two models built on the same underlying system for demanding technical work. Fable includes cyber and biology safeguards and is available through Claude platforms, AWS, Google Cloud and Microsoft Azure, while vetted cybersecurity and life-sciences organizations can apply for Mythos through verification programs.…
Anthropic released Claude Fable 5.2 and Claude Mythos 5.2, describing them as best suited to ambitious, open-ended technical work. The Verge reported that Fable 5.2 was available immediately. Anthropic set pricing at $8 per million input tokens and $40 per million output tokens. During prerelease testing, Anthropic says the models formalized a proof of Fermat's Last Theorem and solved two open bla…
Meta is deploying specialized models trained with Muse Wonder across Muse and Meta AI, bringing Active Memory and Active Focus to all users of those services. The unnamed models were retrained for consumer-assistant behavior such as deciding when to act autonomously, when to seek confirmation, and how to preserve context after corrections, while Muse Wonder remains available to paid subscribers fo…
Baidu open-sourced ERNIE 5.5 Flash Base, a MoE foundation model with 280 billion total parameters and 14 billion activated parameters, according to AIbase. Baidu says the model was pretrained on 33.5 trillion tokens, supports a 128K pretraining context and up to 1 million tokens during fine-tuning and inference, and uses several attention and routing innovations. AIbase also reports Baidu’s claim …
Tencent Hunyuan released the Hy4 large language model on November 4, according to IT之家 and 36Kr. Compared with Hy4 preview, Hy4 introduces visual capabilities and support for computer operation through graphical interfaces and command lines. Tencent says the model can combine perception, reasoning and tool calling for tasks such as detailed image processing, chart analysis and visual mathematical …
Mistral AI released the open weights for Mistral Large 4, nicknamed Le Chonk, a natively multimodal model with about 1 trillion total parameters and 52 billion active parameters. Mistral says the model delivers strong agentic coding and defensive-security performance, with three reasoning modes and support for self-hosted deployment under a modified MIT-style license. Independent coverage reports …
Google DeepMind announced Gemini 4 Carbon, a frontier model focused on coding, agentic workflows, long-horizon tasks and cybersecurity defense. Google reported a score of 59 and first place on the Artificial Analysis Intelligence Index v5.0, along with 88.2% on SEC-bench Pro and a 0.0% attack success rate on Gray Swan's Indirect Prompt Injection benchmark. The model is initially limited to selecte…