Axolotl is a free, open-source tool for fine-tuning large language models. It brings techniques such as FSDP with qLoRA, LoRA+, Multipack and PEFT/LoRA into one configurable workflow driven by YAML files, and supports a wide set of architectures from Hugging Face transformers, including Llama, Mistral, Qwen, Gemma, Falcon, Phi, RWKV and GPT-OSS. The project says new updates arrive every few weeks and claims fine-tuning that is sometimes 3 to 5 times faster than elsewhere.
Machine learning engineers, research labs, model builders and enterprise AI teams are the users. People new to fine-tuning can start from existing YAML configurations in the repository rather than tuning every setting, while advanced teams compose techniques on new model structures. Organizations with compliance needs run Axolotl in their own cloud with Docker and Kubernetes and keep their data in-house rather than uploading it to an outside service. The community includes more than 170 contributors and 500-plus active Discord members who answer questions. Beginners are pointed to a getting-started guide and community recipes, so early runs do not require deep configuration knowledge.
Common projects include fine-tuning a 120B GPT-OSS model across multiple GPUs, training a 405B model on Lambda, and building process reward models. Cookbooks and notebooks supply ready recipes, including a playful example that makes a model talk like a pirate. Users run it on providers such as Runpod, Latitude, Modal and Jarvislabs, with recent updates adding GRPO support and direct pushes to Modal. The framework is used by research companies, model builders and platforms including Nous Research, Modal, Replicate and OpenPipe.
