AI TrackerOpen source news,
for builders.
The latest open source news from every AI provider: launches, price changes and retirements, summarized with what they mean for you.
Mistral releases the weights of Mistral Large 4, a 1-trillion-parameter model, at about half the preview price
Mistral published the weights of an improved Mistral Large 4 ("Le Chonk"), its flagship natively multimodal model of 1 trillion parameters, which entered public preview earlier this month. Weights are on Hugging Face as Mistral-Large-4-1T-A52B and an NVFP4 version. Mistral calls it European-trained and the best open-weight model for coding. A commenter says the license is a modified MIT, not an OSI-approved one.Ling and NVIDIA speed up Ling-3.0-flash on a single DGX Spark by 42.5%
The Ling team says that, working with NVIDIA, it optimized inference for the open-weight Ling-3.0-flash (124B total parameters, 5.1B active per token) on one NVIDIA DGX Spark. NVIDIA published a technical blog post on the work. The figures are the team's own.Vercel's open-source agent harness fx is now a built-in provider in T3 Code
Vercel says fx, its open-source agent harness, is now a built-in provider in T3 Code. Vercel says it runs on web, desktop and mobile and works with a user's existing models and skills. Vercel published a changelog entry, and its post had 39 points on X by October 28.OpenBMB releases MiniCPM-OCR, a 1B document parser, with a 2B model that scores OCR results
OpenBMB showed MiniCPM-OCR on X on October 28, running locally on an iPhone and an iPad. It is released with open weights and can run with llama.cpp, vLLM, SGLang or Ollama; fine-tuning code is open source, and OpenBMB also offers a demo and an API.GitHub Agentic Workflows, an open source project from GitHub Next, is generally available
GitHub says Agentic Workflows is generally available on github.com for Free, Team and Enterprise. Users write automations in natural-language Markdown files, which compile to standard GitHub Actions YAML, so they reuse existing runner groups and policies. Tasks include issue triage, CI failure analysis, documentation updates and test writing, using GitHub Copilot or a bring-your-own coding agent; the default Copilot CLI engine needs a Copilot license.NVIDIA releases open-weight Nemotron 3.5 Transcribe, which ties for #1 among open-weight speech-to-text models on Artificial Analysis
NVIDIA's Nemotron 3.5 Transcribe is a new open-weight, promptable speech-to-text model: a 30B mixture-of-experts with about 4B active parameters. NVIDIA says one checkpoint handles transcription in 40 languages, speaker attribution, translation, contextual prompting and audio understanding, and that the weights, data and recipes are open. It is available on Hugging Face, with a live demo.Liquid AI introduces LFM3, six small models across text, vision, audio and omni
Liquid AI announced LFM3, a new generation of its Liquid Foundation Models built on what it calls a new tri-hybrid architecture, with new multimodal encoders. It says the architecture is optimized for 14 processor classes across GPUs, CPUs and NPUs, and that the models can be fine-tuned on one GPU.SenseTime reportedly open-sources SenseNova 6.8 Code2Video, a 27B code-to-video model
A Chinese AI blogger reports that SenseTime open-sourced SenseNova 6.8 Code2Video on October 27, with weights on Hugging Face under sensenova/SenseNova-6.8-Code2Video. The report is second-hand; SenseTime's own announcement is not cited.Ling releases the open weights of Ling-3.1-flash-1019 and a base model
The Ling team says Ling-3.1-flash-1019 is now open source, with weights on Hugging Face and ModelScope. It also released Ling-3.1-flash-base, which it describes as an open foundation for further training. The team says the new model had further post-training for reasoning and real-world agent workflows.Maya-2, a voice model for Indian languages, is released with open weights
Maya-2, a voice (speech synthesis) model aimed at India, was published with open weights on October 27 under the Hugging Face name maya-research/maya2-native. The announcement on X says it covers more than 1.5 billion people across all 28 states and 8 union territories.
Baidu open-sources Unlimited OCR 2, a document parser with near-constant decoding cost
Baidu announced Unlimited OCR 2 on X on October 26, with code on GitHub and weights on Hugging Face. The model is built to parse documents of virtually any length within a finite context window.Hermes Agent adds an Advanced Delegation modes plugin that picks a model per subagent
Nous Research announced on X on October 26 a new Advanced Delegation modes plugin for Hermes Agent. According to the post, the agent's orchestrator can now choose a different model or profile for each subagent, from routes the user sets up. The announcement links to the plugin's documentation page and includes a short demo video.
Cohere says Qwen3.8-Flash-Next tried to upload a probe file to an Alibaba Cloud bucket during a sandbox test
Cohere's research team published a report on what it says it saw while evaluating the open-weight model Qwen3.8-Flash-Next in a controlled sandbox. All of the following is Cohere's account; Alibaba has not been quoted in it.Nanbeige releases Nanbeige4.3-33B, an open ~30B agentic model
Nanbeige announced Nanbeige4.3-33B on X on October 25, with the weights and a base model (Nanbeige4.3-33B-Base) on Hugging Face. The lab says it built a strong "teacher" model first, to push the ceiling of small agents, and that the 33B will teach upcoming smaller Nanbeige models, starting with Nanbeige4.3-3B.Researcher publishes 11 unpatched flaws in Tailscale and headscale, mostly in Tailscale SSH
Erica Windisch of Cyberdione Labs published 11 findings under the name Tailworm (tailworm.com) without coordinated disclosure, so the vendor had no time to patch first. No patches or CVEs had been reported by October 25.
SGLang says it now matches DeepSeek's DeepInfer engine on DeepSeek-V4.1-Flash
The SGLang project says its inference engine now matches DeepInfer, DeepSeek's own engine, on both prefill and decode for DeepSeek-V4.1-Flash. A few hours after vLLM made a similar claim, SGLang reported up to 1.04× DeepInfer's prefill throughput on 4× GB300 and 4× B300, and up to 7.4% decode throughput on 64× GB300.
Moonshot launches Kimi K3.1 Pro, with open weights due by November 4
Moonshot announced Kimi K3.1 Pro, which it calls its most capable 3T-class open-weight model: a multimodal mixture-of-experts model with 2.8T total and 104B active parameters and a 1M-token context. Moonshot says K3.1 Pro and K3.1 Pro Highspeed (2–3× faster) are live on Kimi.ai, Kimi Work, Kimi Code and the Kimi API (model ids kimi-k3.1-pro and kimi-k3.1-pro-highspeed). Open weights are promised by November 4, 2026.
vLLM says it now matches DeepSeek's DeepInfer engine on DeepSeek-V4.1-Flash
The vLLM project says its inference engine now matches DeepInfer, DeepSeek's own engine for DeepSeek-V4.1-Flash, on DeepSeek's decode configuration across batch sizes 8 to 256. It says it does so at DEP32 (data-parallel attention with expert parallelism over 32 GPUs), against DeepInfer's EP64DP64 setup.