$ The best sources for open-weight model news

Where to follow new open-weight model releases: lab announcements, the Hugging Face hub, inference engine release notes and analysts, and how to tell which releases you can actually run.

$ subscribe --free
Get this checkup by email each morning: choose your own sources, double opt-in, one-click unsubscribe. Or follow the RSS feed.
Choose sources & subscribe →

What counts as an open-weight model

An open-weight model is one whose trained weights you can download and run on your own hardware. That does not automatically make it open source: many models come with their own licence that limits commercial use or requires attribution, and only a few projects also publish their training data and code. Read the licence on the model card before you build on a model.

New releases usually reach you in three steps. A lab announces the model, the weights appear on a hub such as Hugging Face, and a few days later the inference engines and runners add support, often together with smaller quantized versions that fit on consumer GPUs.

Lab announcements

The labs that publish open weights announce them on their own blogs. Mistral AI's news page covers its open-weight and hosted models, and the Hugging Face blog regularly explains new model families together with the people who trained them. Fully open projects such as Ai2's OLMo also publish their training data and code, which makes them a good reference for how a model was built.

The Hugging Face hub

Almost every open-weight model ends up on Hugging Face, so its blog and the community-curated daily papers list are the closest thing to a single place for new releases and the research behind them. The release notes of the Transformers library show when a new architecture can be loaded with standard tools.

When can you run it?

A model is only useful to you once your tools support it. The release notes of llama.cpp, Ollama and LM Studio show when a new architecture or quantization format works on a normal desktop or laptop, and vLLM's release notes do the same for GPU servers. Fine-tuning tools such as Unsloth publish support for new models quickly, which is a good sign that a release is getting real use.

Analysis and comparisons

Benchmark tables in announcements are written by the people who made the model. Interconnects by Nathan Lambert and Ahead of AI by Sebastian Raschka explain training and post-training methods in depth and put new releases in context, which helps you decide whether a model is worth trying. For general coverage, Ars Technica reports on model releases and the debate around them.

How to combine them

A practical setup is the Hugging Face blog for new releases, the release notes of the one runner you use, and one analyst for context. Try a new model on your own tasks before you switch, because results on public benchmarks often do not carry over to real work.

If you would rather not check these pages every day, Linux Doctor sends a free daily email with the new items from these sources and publishes the same list on the latest checkup page.

All sources mentioned

Hugging Face Blog: Open models, datasets, tooling and tutorials from the open-source ML hub.
Hugging Face Daily Papers: Community-curated trending ML papers (community-run feed).
llama.cpp Releases: Release notes for llama.cpp, the core local-inference engine.
Ollama Releases: Release notes for Ollama, the easy local model runner.
LM Studio Blog: News from LM Studio on running models on your own machine.
Mistral AI News: Announcements for Mistral's open-weight and frontier models.
Interconnects: Nathan Lambert on open models, post-training and RLHF.
Ahead of AI (Sebastian Raschka): Hands-on explanations of LLM research and training techniques.
Ars Technica AI: Tech journalism on AI products, models and industry.
arXiv cs.CL: New papers on language models (high volume).
Inco AI Blog (Decode): Decode, the research and announcement blog of the inference company Inco AI: speculative decoding (DFlash parallel drafters), day-0 support and quantized checkpoints for open models like GLM, Kimi and Qwen, and Splash, their open-source local inference engine for Apple silicon. Low volume, a few posts a month.
Ai2 News (OLMo, Tülu): News from the Allen Institute for AI, which publishes fully open models such as OLMo and Tülu together with their training data and code.
Hugging Face Transformers Releases: Release notes for Transformers: shows when a new model architecture can be loaded with standard tools.
Unsloth Releases: Release notes for Unsloth, the fast fine-tuning library that adds support for new open models within days.
Ollama Blog: Announcements from Ollama: newly supported models and features for running them locally.
SGLang Releases: Release notes for SGLang, a fast inference server for LLMs on Linux GPU servers, often with day-0 support for new open models.

Open models news latest checkup all sources