Open weights

Open weights is narrower than open source: the parameters are published, but training data and full training code usually are not, and licenses range from permissive Apache-style terms to community licenses with usage restrictions. Meta's Llama releases, Mistral's models, and DeepSeek's R1 and V-series made open weights a serious alternative at ever-higher capability levels.

Economically, open weights decouple the model from any single vendor. There is no per-token fee to the creator; you pay whoever serves it — a competitive market of inference providers — or your own hardware bill if you self-host. Competition among hosts drives the price of popular open models well below comparable proprietary rates.

The tradeoffs are operational: quality varies across hosts depending on quantization and serving choices, and self-hosting trades API convenience for infrastructure work. For high-volume, privacy-sensitive, or fine-tuning-heavy workloads, open weights are frequently the cheapest viable path.