Open weights
Model weights publicly available for download and self-hosting
What is open weights?
Open weights means the model's parameters are publicly available — you can download, run, fine-tune, and inspect the model yourself. Llama 3.1 405B from Meta, Mistral 7B, and DeepSeek R1 are all open weights. The practical impact: zero per-token cost if you self-host. The tradeoff: you pay for compute infrastructure instead. At scale (millions of requests/month), self-hosting open weights models is often cheaper than API access.
Why it matters
At $0.00 per token, open weights models have infinite Value Density — limited only by your infrastructure cost. A Llama 3.1 70B model running on a $1/hour GPU processes roughly 50 tokens per second. That is $0.02 per million tokens in compute cost — 150× cheaper than the same model via a cloud API. sourc.dev tracks which models have open weights and which inference providers host them.
Where models stand
10 models with open weights enabled:
Data available for 11 of 271 tracked entities. Last updated 2026-03-30.
How sourc.dev tracks this
sourc.dev verifies open weights manually from official provider documentation, API responses, and published specifications. Every data point includes a source URL and verification date. When a value changes, the old value is preserved in the history table and the new value is recorded alongside it. Nothing is overwritten — the full timeline is always available.
Open weights means the model parameters are publicly available for download and use. Open source, in the traditional software sense, also requires access to training data, training code, and a permissive licence. Most "open" models are open weights, not fully open source.
Generally yes, but it depends on the licence. Models like Llama have licences that permit fine-tuning with some restrictions. Others like Mistral models use Apache 2.0 with fewer restrictions. Always check the specific licence.
The gap has narrowed significantly. Models like Llama 3.1 405B, Mixtral, and DeepSeek V3 compete with closed models on many benchmarks. For specific tasks, fine-tuned open models can outperform general-purpose closed models.