About
Open weights move fast — a frontier model can appear overnight and reset the field by morning. The Open Weights exists to make that legible: every release, deduplicated and explained, with the links you actually need.
How it works
Every 12 hours, Claude reads Hugging Face, arXiv, GitHub, and the labs' own blogs, then normalizes and deduplicates what it finds — so one release is one entry, never a dozen.
Gemini drafts each story from the structured facts, adding context and a clear 'why it matters' — always citing at least one authoritative source.
A human editor reviews the day's batch before anything major publishes. No unreviewed, mass-generated pages — quality is the whole point.
Editorial policy
We use AI to aggregate and draft, never to publish unreviewed at scale. Every story is built from primary sources, adds original context, and links back to the model card, code, or paper it describes.
Major releases are reviewed by a human editor before publication; the rest flow through a confidence gate. We correct errors openly and update stories as facts change — the modified date on each article reflects that.
Leaderboard data is cached from public sources and refreshed every 12 hours. We link to each source so you can verify methodology yourself.
Hardware estimates
Every release with downloadable weights gets a “Can you run it?” estimate. We start from the parameter count — or the exact weights metadata when Hugging Face publishes it — and work out how much memory the model occupies at each common precision: as published, 8-bit, and 4-bit.
For language and vision-language models we add a KV cache sized for an 8K-token context, read from the model’s own config where possible, then about 1 GB for the runtime itself. Mixture-of-experts models count every expert, because all of them must be resident even though only a few fire per token. Image, video, audio, and embedding models show a single native-precision figure (plus FP8 where it applies) that covers the published checkpoint.
The lowest-memory row sets the tier badge on cards and the “Fits in” filter on the feed. The machine picker treats about 80% of a Mac’s unified memory as usable by the GPU, 90% of system RAM for CPU-only inference, and all of a discrete card’s VRAM.
These are planning numbers, not benchmarks. A given quantized build can be a little larger or smaller, long contexts and batching add memory, and models that ship only in a reduced precision never show rows above it.