Axolotl

Axolotl is a free, open-source fine-tuning framework that makes customizing large language models faster, more accessible, and scalable. It supports a wide range of model architectures via Hugging Face Transformers and integrates cutting-edge techniques like LoRA, FSDP+qLoRA, and Multipack to deliver training speeds 3–5x faster than alternatives. Axolotl is used by researchers, model builders, and enterprise AI platforms who need full control over their data and infrastructure.

• Broad LLM Support — Fine-tune popular models including LLaMA, Mistral, Falcon, Gemma, Qwen, Cerebras, MS Phi, RWKV, and more out of the box.
• Multi-GPU Training — Integrates with xformers and DeepSpeed for efficient distributed training across multiple GPUs.
• YAML-Based Configuration — Get started quickly using pre-built YAML configs in the OSS repo without needing to write custom training code.
• Advanced Fine-Tuning Techniques — Supports FSDP+qLoRA, LoRA+, Multipack, PEFT/LoRA, GRPO, and Process Reward Models for state-of-the-art results.
• Bring Your Own Data (BYOD) — Keep your data on-premises or in your own cloud — no need to upload to third-party services, ensuring compliance and data governance.
• Flexible Deployment — Run anywhere: your own cloud, Docker, Kubernetes, or partner platforms like Runpod, Modal, Lambda Labs, and Jarvislabs.
• Active Open-Source Community — Over 170 contributors and 500+ active Discord members provide support, recipes, and cookbooks to help you get started fast.
• Axolotl Cookbooks — Pre-made recipes and notebooks (e.g., fine-tuning GPT-OSS 120B, talk-like-a-pirate examples) make experimentation easy.
• Enterprise-Ready — Supports custom Docker and k8s setups, full data control, and is trusted by research companies and Gen AI platforms at scale.