Axolotl is a free, open-source framework for post-training and fine-tuning large language models, created by Wing Lian and maintained by Axolotl AI under the Apache 2.0 license.
The framework supports full fine-tuning, LoRA, and QLoRA (including LoRA+ and QLoRA with FSDP), as well as preference and reinforcement learning methods: DPO, ORPO, KTO, GRPO, and process reward modeling. Configuration is driven by a single YAML file, with pre-configured example recipes for common setups. Performance features include Flash Attention variants, multipack sample packing, sequence parallelism for long-context training, ScatterMoE kernels, and multi-GPU and multi-node training via FSDP and DeepSpeed; the project claims fine-tuning 3-5x faster than alternatives.
Axolotl supports models compatible with Hugging Face Transformers, including GPT-OSS, Llama, Mistral, Mixtral, Falcon, Gemma, Qwen, Pythia, and RWKV, plus multimodal models such as Llama-Vision, Qwen2-VL, and Pixtral, and audio models like Voxtral.
There is no proprietary cloud: Axolotl runs self-hosted via Docker or Kubernetes and integrates with GPU providers including RunPod, Lambda Labs, Modal, and Baseten. Documentation is available at docs.axolotl.ai. The project is completely free with no paid tiers, and its GitHub repository has around 12,000 stars.