Model announcements

With new models released several times a month, staying truly up to date is challenging. By producing curated model announcements, our team and your team can stay up to date.

We aim to produce an organized, TLDR-style summary that positions the model in the context of the broader open-weight space. We focus on what the important and practical implications of the model release are.

When we publish

Model announcements often happen in advance of when the model weights are published. Sometimes the full technical report comes out with the announcement, other times it comes out with the weights.

We aim to publish shortly after the technical report is out. If we publish before the weights are public, any content relying on the weights (for example, our own benchmark) can be updated once the weights are public.

How we stay informed

  • Hugging Face is the source of truth for when new model weights are live.
  • OpenRouter is a good proxy for when a new model is announced or gets API access. It's almost always a day-zero partner.
  • Both have free public APIs. Only a few lab blogs have an RSS feed.

The model radar checks both every 15 minutes and posts to #model-radar in Slack when a lab on our watch list publishes a new model.

To watch a new lab, add a line to watchlist.json in a pull request. The lab's model page on OpenRouter links to its Hugging Face model, which gives you both names.

{ "lab": "Tencent", "huggingface": ["tencent"], "openrouter": ["tencent"] }

We also follow the labs' X accounts by hand to catch early chatter or a "coming soon" post.

What goes in a post

  • How long since the last model release from this lab or model family?
  • What differentiates it from its predecessors? Is it bigger, smaller, specialized?
  • What the announcement and technical report emphasize.
  • What kinds of work or industries this model might be relevant to: who should pay attention?
  • Interesting or important changes, such as an update to the attention architecture or training strategy.

Also anything handy:

  • Licensing details or updates: can you use it commercially, are there revenue caps, etc.
  • Links to the announcement, the weights, the technical report and the model on Hugging Face.
  • Relevant benchmarks.

Gradiated benchmarks

Coming soon.