LLaMA
The default choice when you need weights you can actually hold, fine-tune and run offline, with an ecosystem of quantisations and inference stacks already built around it. The licence is not OSI-open, and Meta has been reorganising the Llama web presence, so links rot faster than the models do.
Pros
- Weights are downloadable, not just API-accessible
- Broad tooling support: llama.cpp, vLLM, Hugging Face and most inference runtimes
- Multiple sizes so you can trade quality against a single GPU
- Commercial use permitted for almost everyone under the MAU threshold
Cons
- Community licence is not open source; attribution and the MAU cap both bind
- Download links are pre-signed and expire after 24 hours or five uses
- Meta keeps moving the site: llama.meta.com now forwards twice before landing
- Running it is your cost and your operational problem