Models whose trained parameters are publicly downloadable — self-hostable, fine-tunable, and servable by any inference provider, unlike closed API-only models.
An open-weight model publishes its trained parameters (usually on Hugging Face), so anyone can download, run, fine-tune or re-serve it. Closed models (GPT, Claude, Gemini) are only reachable through their maker's API.
Open weights bring price competition — dozens of providers serve the same model, driving costs down — plus data control (run it in your own VPC), customization via fine-tuning, and immunity to deprecation. The trade-off has historically been capability, but top open models now sit within a few points of frontier closed models on intelligence benchmarks.
"Open weights" isn't the same as "open source": many licenses restrict commercial use or require attribution. Check the specific license before building on one.