Lemur

Why we included this project

Lemur is a 70B open-weight model that pairs natural-language reasoning with solid coding ability, so it works as a single backbone for teams building tool-using agents rather than a chatbot or a ready-made agent framework. You load the base and instruction-tuned checkpoints through the standard transformers pipeline and wire them into your own workflows. The repo documents the two-stage training recipe, continued pretraining of Llama-2-70B on a code-heavy corpus followed by instruction tuning on text and code examples, which makes it a useful reference if you are studying how to balance language and coding skills in one model. The weights are Apache-licensed and served with vLLM, so self-hosting for evaluation or fine-tuning is straightforward. Teams looking for open alternatives to commercial models for agent-style tasks will find the checkpoints and the evaluation scripts for benchmarks such as MMLU, HumanEval, and WebArena ready to use.