SoTA Feed

At a glance

SoTA Feed tracks open-weight conversational model releases from a selected set of labs. It is worth adding to a feed reader if you want to spot new model files by format and download size before reading the full model cards.

Good fit

People following downloadable models from labs such as NVIDIA and Moonshot AI, especially when they have a storage budget.

Consider something else when

Skip it if you need a complete registry, benchmarks or a reliable estimate of the RAM or GPU memory a model needs during inference.

Start here

Open the site and set a total size limit that fits your storage. Open one matching release and confirm its format and license with the original publisher. Use the headline-only feed if you just want a release list.

Cost and requirements

The public site and RSS feed worked without an account. The site lists a $3 monthly option for an ad-free site and personal feed.

Limits to know

Download size and active-parameter estimates are screening aids. A mixture-of-experts model can activate only part of its parameters per token while still needing access to the rest of its weights. Check the runtime’s actual memory requirements separately.

Related work

• NVIDIA, whose model releases appear in the feed (context)
• Moonshot AI, whose Kimi releases appear in the feed (context)