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.
Asahi Linux moves all its code to the GPL, citing AI-assisted code laundering
Asahi Linux, the project that brings Linux to Apple Silicon Macs, has announced in a progress report that it is moving all of its repositories from the MIT license to the GNU General Public License version 2 or later, so that its code will be under copyleft licenses only. The project announced the change a few days before November 9.
Nous Research releases Arete-27B, Qwen3.8-27B post-trained to cut a "party line" rate from 22% to 1%
Nous Research announced Arete-27B on X on November 6. It is Qwen3.8-27B post-trained, in the lab's words, "to reinforce Western values". Nous argues that the strongest open models are Chinese and ship with state-aligned behavior in their weights.Harness Leaderboard compares nine coding-agent harnesses on open models; a bash-only agent tops it
Joel Niklaus released the Harness Leaderboard, which compares nine coding-agent harnesses running open models. In each comparison only the harness changes: the model, provider and reasoning effort stay fixed, and every task gets three attempts. Niklaus says he ran 28,849 trials for $4,746 of inference, using GLM-5.3-Flash, DeepSeek-V4.1-Flash and Flint on Terminal-Bench 2.1 and DeepSWE. The post had 625 points on X by November 6.Village: open-weight models are the best buy at 3% of budgets, down from 41% in June
Village, which publishes the LLM Atlas, reports that closed models now hold almost every point on the price-performance frontier. By its estimates, an open-weight model is the strongest model at 3% of budgets, nearly all at the cheap end (Google's Gemma 4 E4B and Xiaomi's MiMo-V2.6-Flash). In June the share was 41%, when DeepSeek V4 Flash alone led nearly a third of budgets. Village says today's share is the lowest it has measured; at each quarter's end since 2024 it was at least 6%.
Artificial Analysis ranks FLUX 3 Image Klein #2 among open-weight text-to-image models
Artificial Analysis says Black Forest Labs' FLUX 3 Image Klein, a 5.6B-parameter open-weight image model released on October 15, is the #2 open-weight model on its AA-Image-T2I v2.1 leaderboard, neck and neck with #1 Qwen-Image-2.1. On AA-Image-Editing v2.1 it ranks #3 among open-weight models. The rankings are Artificial Analysis's own.Moonshot releases the open weights of Kimi K3.1 Pro
Moonshot posted on X that it is releasing the model weights of Kimi K3.1 Pro, on Hugging Face as moonshotai/Kimi-K3.1-Pro, and said "we'll keep advancing open intelligence." It had promised the weights by November 4 when it launched the model on October 23.Hermes Agent v0.23.0 adds a browser desktop, /rewind, group chats and spend limits
Nous Research released Hermes Agent v0.23.0, which it calls the Hearth Release. Its highlights, as the company lists them:Artificial Analysis scores open-weight Nanbeige4.3-33B 20 on its Intelligence Index, level with Qwen3.8 27B
Artificial Analysis says Nanbeige4.3-33B, the open-weight model from Nanbeige (the foundation-model team at the Chinese hiring platform BOSS Zhipin), scores 20 on its Intelligence Index v5, the same as Qwen3.8 27B (xhigh). Both are dense models, with 33B and 27B total parameters. Artificial Analysis says Nanbeige's model has stronger terminal use but weaker knowledge, and sets a new Pareto-optimal point for Intelligence Index against total parameters.Baidu open-sources ERNIE-5.5-Flash-Base, a 280B-parameter sparse MoE
Baidu has published ERNIE-5.5-Flash-Base on Hugging Face. Its model card describes ERNIE 5.5 Flash as an efficient sparse mixture-of-experts language model with 280B total and 14B active parameters, pretrained from scratch on 33.5T tokens with a 128K training context. Fine-tuning and inference support up to 1M tokens of context.Ricursive AI releases Ripple-65B Preview, a 65B dense open-weight model for agentic coding
Ricursive AI released Ripple-65B Preview on November 4, a 65-billion-parameter dense language model it says it pretrained from scratch. The company claims it sets a new state of the art for U.S. open-weight models on agentic coding tasks, ahead of much larger models it names: Muse Spark Open-816B, Beam-501B, Inkling-975B and Nemotron 3 Ultra-550B. These are the company's own claims.Nous Research puts Hermes Agent's plugins, connectors and skills in one marketplace
Nous Research said on X on November 4 that plugins, connectors and skills for Hermes Agent are now unified in a single Hermes Marketplace, where users can browse and add extensions in a few clicks. The marketplace is at hermes-agent.nousresearch.com/marketplace.Cloudflare open-sources a security-audit skill for coding agents
Cloudflare's developer account released cloudflare/security-audit-skill, an open-source skill that turns a coding agent into a team of security auditors. Cloudflare says the skill seeded its own vulnerability-discovery harness.Perplexity releases pplx-embed-v2-context, open-weight vision-language embedding models in 9B and 0.6B
Perplexity released pplx-embed-v2-context, two open-weight vision-language embedding models (9B and 0.6B) on Hugging Face, with a paper written with turbopuffer. Both return one int8 vector per document, or one per chunk that sees the whole document, and both also run in a non-contextual mode with one embedding per document and no chunking.
OpenCode is rebuilt to run inside browser devtools
Adam (adamdotdev) said on X on November 2 that he rebuilt OpenCode to run inside a browser's devtools console, using console.log statements for its interface. It is built on a framework he links at textjs.org, and he invites developers and their agents to build other console apps with it. A note with the post says a VS Code extension is in the works and a Chrome devtools extension is still to come.
Anthropic releases FlashProve, an open-source prototype for proving which model produced a response
Anthropic published research on provable inference: a cryptographic proof (a zkSNARK) that a response came from a specific committed set of model weights, checkable by a user who does not have the weights. The post says it could help detect a model silently swapped by a routing service or tampered with by an attacker, and could give evaluators and regulators a way to check what is being served.
Hydrogen-1.0, an open-weight scientific model post-trained from GLM-5.3, is released
Chenru Duan announces Hydrogen-1.0, which the team calls its first open-weight model for scientific research. It is GLM-5.3 (from Z.ai) post-trained on the team's internal scientific data, lab records and discovery trajectories.
MiniMax lists community quantizations for running M3.1 Flash locally
MiniMax posted a roundup of community-made quantized versions of its 428B-parameter open-weight M3.1 Flash, aimed at people running it on their own hardware. MiniMax says lower-bit weights cut the memory needed to load the model and the weight data moved during decoding, and that making this useful takes careful calibration, efficient low-bit kernels and placing experts across GPU and system RAM.