AI news for Nov 15, 2026
12 stories
- ResearchOpenAIOpenAI explores a lattice definition of the Standard Model with GPT-6 Astra
OpenAI published an exploration of a candidate nonperturbative lattice definition of the Standard Model, and of the role its GPT-6 Astra model played in developing the idea into a detailed proof. The page presents the lattice formulation as a step toward defining the Standard Model on a lattice and describes how the model contributed to that work.
- BusinessHugging FaceHugging Face moves Inference Providers routing onto OpenRouter
Hugging Face has replaced its own router for language model inference with OpenRouter's, as the first step of a partnership between the two companies. Requests made with a Hugging Face token now go through OpenRouter, which picks among 80+ providers based on live price, speed and reliability and retries a failed call on another provider. A suffix on the model id or account settings can still pin or block providers. The router supports OpenAI's Chat Completions and Responses APIs and Anthropic's Messages API. Pricing is unchanged, with no markup added and $2 monthly credits for PRO users. Custom provider API keys no longer work for these models, a breaking change; image, video and audio generation stay on Hugging Face's own inference.
- ModelsHugging Face PapersSage publishes technical report on 320B-parameter S1.5 Flash model
Sage has released a technical report describing Sage S1.5 Flash, a 320-billion-parameter mixture-of-experts model built to reason over scientific evidence and act in interactive environments. Its Sci-Perception component uses dedicated encoders for molecular graphs and biological sequences and treats spectra as ordered patches of measurements; projectors map those representations into a shared backbone that also handles text and images. Agent training used verifiable environments for tool use, search and coding, plus workflows for machine-learning experiments and scientific software, with completion checked against execution results or submitted artifacts. The report benchmarks the model on 33 tasks across seven domains, where it ranks first on OfficeQA Pro, MLE-bench Lite, Harvey LAB and SpreadsheetBench 2, and second on OSWorld-Verified, WebArena-Verified and DeepResearch Bench II.
- Open sourceHugging Face PapersK2-Horizon Releases Six Fully Open Language Models, Including a 375B MoE
K2-Horizon is a family of six openly released language models, ranging from a 0.9B dense model to a 375B-A23B mixture-of-experts model, built for reasoning and agentic work in math, science, software engineering, deep research and office automation. The authors say the models lead or are competitive with strong open models of their size on agentic and coding benchmarks. Two attention changes, Mixture-of-Value Attention and softplus attention gating, let a 36B-A4B model nearly match a 32B dense model trained on the same data with one-seventh the active parameters. A companion release, Uno, provides diffusion adapters that speed decoding by up to 2.2x without a draft model. Checkpoints from every training stage, pre-training data, logs, and training and evaluation code are published, along with the data-building pipelines.
- ResearchHugging Face PapersAuthors Release InclusionMed Benchmark for Inclusive Medical AI
The team behind InclusionMed, an open medical AI evaluation initiative, published a benchmark suite and a model index. The InclusionMed Atlas organizes evaluations by three healthcare roles and three AI capabilities; its first release selected four public benchmarks from 388 candidates after structured audits and independent reproduction. Adding the private diagnostic reasoning set InclusionMed-DxFiles, five benchmarks now cover three of the nine role-capability cells. Frontier language models scored between 17.5 and 49.1 out of 100 on the aggregate InclusionMed Index. A model's rank changed with the role or capability examined, and the ten top-ranked models all scored higher on patient communication than on clinical decision-making. Models with similar Index scores differed substantially in estimated inference cost.
- ProductsBrowser UseBrowser Use makes BU 4.0 the default model in browser-use 1.0.6
Browser Use published version 1.0.6 of its browser-use library, and the agent now defaults to BU 4.0 instead of bu-fast. Agents created with Agent(task) or ChatBrowserUse() call bu-4.0 through the Browser Use gateway, and developers can pick bu-4.0-mini for simpler work or bu-4.0-pro for harder tasks. Running with no llm setting costs more: BU 4.0 is priced at $1.80 per million input tokens and $9.00 per million output tokens, against $0.12 and $0.60 for bu-fast, and the default runs at high rather than medium reasoning. Passing llm='bu-fast' keeps the earlier model and price. The release also opens a Browser Use login on the first run without a key and prints live-view links for cloud browsers.
- DevelopersLangChainLangChain ships langchain-mistralai 1.2.0 with reasoning-model streaming fix
LangChain published version 1.2.0 of its langchain-mistralai partner package on GitHub, following 1.1.6. The release fixes streaming behavior for reasoning models and routes the ChatMistralAI chat class through the LangSmith gateway. It also bumps dependencies shared with the Groq and xAI integrations and refreshes the bundled model profile data. Separately, end-of-life Python 3.10 is dropped from the partner packages, so the integration now targets newer Python versions.
- DevelopersLangChainLangChain ships langchain-groq 1.2.0 with LangSmith gateway routing
LangChain published version 1.2.0 of langchain-groq, the Python partner package that connects its framework to Groq's inference API. According to the changelog, the release routes ChatGroq calls through the LangSmith gateway and removes end-of-life Python 3.10 support from the repo's partner packages. The version also refreshes model profile data, relaxes upper bounds on third-party dependencies, and includes several dependency bumps and test fixes.
- Open sourceLangChainLangChain ships langchain-xai 1.4.0 with LangSmith gateway routing
LangChain published version 1.4.0 of langchain-xai, its partner package for xAI models. The release routes the ChatXAI chat model through the LangSmith gateway so requests from that integration can pass through LangChain's gateway layer. It also removes Python 3.10, which the project marks end-of-life, from its partner packages, refreshes bundled model profile data, relaxes upper bounds on third-party dependencies, and bumps libraries including aiohttp, certifi, anyio and urllib3. The structured outputs guide now points to xAI's official documentation. The package is available for developers building LangChain applications against xAI models; this is a feature and maintenance update rather than a new model.
- DevelopersLangChainLangChain ships langchain-deepseek 1.2.0 with LangSmith gateway routing
LangChain published version 1.2.0 of langchain-deepseek, the partner package that connects its Python framework to DeepSeek models. The release routes the ChatDeepSeek wrapper through the LangSmith gateway, so requests from that class pass through LangChain's gateway layer. The same update set drops Python 3.10, which has reached end of life, from LangChain's partner packages, and relaxes upper bounds on third-party dependencies. It also refreshes stored model profile data and bumps the certifi and urllib3 dependencies.
- ModelsVercelByteDance's Seedance 2.5 Flash video model arrives on Vercel AI Gateway
Vercel added ByteDance's Seedance 2.5 Flash to AI Gateway, a faster and lower-cost alternative to Seedance 2.5. According to Vercel, its output comes close to the standard model on human realism, reference-based editing and prompt adherence. The model produces clips with synchronized audio from text prompts or from image, video and audio references. Clips run 4 to 30 seconds at up to 720p and 24 frames per second in landscape, square or portrait formats; users can also edit or extend existing videos or supply first and last frame images. Developers call it through the model ID bytedance/seedance-2.5-flash with AI SDK 7, and it is also available in Vercel's model playground.
- ModelsVercelAlibaba's Qwen4-Max reasoning model arrives on Vercel AI Gateway
Alibaba's Qwen4-Max, the first model in the Qwen4 series, is now callable through Vercel's AI Gateway as alibaba/qwen4-max. Vercel describes it as a reasoning model with native vision and a 1M-token context window, aimed at coding, professional work and long-running agent tasks. It can build scenes in Blender and games in Unity, Unreal Engine and Godot, then use visual feedback to correct camera, animation and lighting problems; for finance, legal and healthcare work it calls MCP tools across large document sets and produces sourced spreadsheets, decks and reports. Alibaba is open-sourcing a Live Delivery plugin for its draft-then-refine workflow. Developers can pick the model in Claude Code, Codex and Cursor after installing the latest Vercel CLI.