The fine-tuner's default toolkit
Axolotl began as an open-source project by Wing Lian and grew into the de facto community standard for post-training large language models. The company Axolotl AI, based in San Francisco, now stewards the project, which counts 170+ contributors and powers fine-tunes from research collectives like Nous Research and platforms like Modal and Replicate.
How it works
Axolotl wraps the messy landscape of fine-tuning techniques — full fine-tuning, LoRA, QLoRA, LoRA+, FSDP+QLoRA, DPO, RLHF-style preference tuning, sequence parallelism and Multipack sample packing — behind declarative YAML recipes. It supports the major open model families (Llama, Mistral, Qwen, Gemma, Falcon and more), integrates DeepSpeed and xformers for multi-GPU efficiency, and claims 3-5x faster training than naive setups. It runs anywhere: local GPUs, Docker, Kubernetes or cloud platforms, with users keeping full control of their data.
Why it matters
Most notable open fine-tuned models of the past few years were trained with Axolotl. For teams that want owned, specialized models without building training infrastructure, it remains the fastest proven path.