OpenLLaMA 13B

OpenLLaMA 13B is the largest first-version model from OpenLM Research's effort to produce a permissively licensed open-source reproduction of Meta's LLaMA. Trained on 1T tokens, it was designed as a drop-in replacement for LLaMA 13B in existing implementations, and both the training framework and the weights are released under the Apache 2.0 license.

The project released 3B, 7B, and 13B models, with the 13B arriving after the smaller sizes. It gave the 2023 open-source community a commercially usable 13B base model at a time when LLaMA itself was restricted to non-commercial research.