What this tracks (and what it doesn't)
This page lists the newest open-weights releases from the labs that matter to local AI — DeepSeek, Qwen, Meta, Mistral, Google's Gemma, OpenAI's gpt-oss line, Microsoft's Phi, NVIDIA, Moonshot, Z.ai, MiniMax, and IBM. It refreshes itself daily from the Hugging Face API; nobody hand-curates it, which is exactly why it's never stale.
Why no Claude, GPT-4-class, or Gemini API models? Honest answer: they're closed — there are no weights to download, so there's nothing to run on your own hardware, which is what this site is about. When a closed lab releases open weights (as OpenAI did with gpt-oss), they show up here automatically.
How to use it
The Released column tells you what's genuinely new; Downloads tells you what the community actually adopted (a two-day-old model with a million downloads is a signal). Sizes are parsed from model names — "(active)" marks Mixture-of-Experts models, where memory follows total size but speed follows the active number. Every sized model gets a one-click Can I run it? link pre-filled with its size.
Filtering, in the open: we keep text-generation and vision-language models, and drop quant re-uploads, base-model duplicates, guard/safety classifiers, and research artifacts. The raw feed is public at /models-feed.json if you want to build on it.