Axolotl
Free, open-source LLM fine-tuning framework driven by one YAML config: LoRA, QLoRA, full tuning, DPO and GRPO, multi-GPU and multi-node
Overview
Axolotl calls itself "a free and open source LLM fine-tuning framework." You describe a whole run in one YAML file (dataset preprocessing, training, evaluation, quantization and inference) and Axolotl handles the rest on top of Hugging Face Transformers and TRL. It covers full fine-tuning, LoRA, QLoRA, GPTQ, quantization-aware training and FP8, preference and RL methods (DPO, IPO, KTO, ORPO, GRPO, reward and process reward models), and vision and audio models such as Qwen2-VL, Pixtral, Gemma 3n and Voxtral. Scaling is its strong side: FSDP2, DeepSpeed, sequence parallelism and multi-node runs through torchrun or Ray. Wing Lian started it in 2023 under the OpenAccess AI Collective; it is now maintained by Axolotl AI, Apache-2.0 licensed, and has no paid hosted product (the company takes sponsorships). Latest release: v0.21.1, October 27, 2026.
The Verdict
Who Should Use Axolotl?
Best For
- Teams fine-tuning on many GPUs or several nodes
- Reproducible runs kept as config files in version control
- Practitioners who want new methods (GRPO, QAT, FP8) without writing trainers
Not Ideal For
- Beginners who want a point-and-click UI (see LlamaFactory or Unsloth Studio)
- Squeezing the most speed out of one consumer GPU (Unsloth is faster there)
- Buyers who need a vendor SLA
What's Great
- One YAML file drives the whole pipeline, so runs are easy to repeat
- Broad method list, from LoRA to GRPO and quantization-aware training
- Strong multi-GPU and multi-node support (FSDP2, DeepSpeed, Ray)
- Used by Nous Research, Modal, Replicate and OpenPipe
Watch Out For
- Errors can surface deep in TRL or Transformers, far from your config
- Slower than Unsloth on single-GPU QLoRA in blog benchmarks
- Frequent breaking changes (FSDP1 was removed in v0.20.0)
Pricing
View all features & details
Key Features
- Full fine-tuning, LoRA, QLoRA, GPTQ
- QAT (int8, int4, FP8, NVFP4, MXFP4) and FP8 mixed precision
- DPO, IPO, KTO, ORPO, GRPO, reward and process reward models
- Vision-language and audio models
- FSDP2, DeepSpeed, sequence parallelism, multi-node via torchrun or Ray
- Datasets from local files, the Hub, S3, Azure, GCP or OCI
Where It Runs
- Docker images and PyPI package
- Integrations with RunPod, Latitude, Modal and Jarvislabs
- Direct pushes to Modal (2026)
Real-World Usage
Community
- Over 170 contributors
- 500+ active Discord members
- Releases every few weeks through 2026
How It Compares
| Feature | Axolotl | Unsloth | LlamaFactory | TRL |
|---|---|---|---|---|
| Interface | YAML config + CLI | Python, Studio web UI | Web UI + CLI | Python + CLI |
| Strongest at | Multi-GPU, multi-node | Single-GPU speed and memory | Ease of start, model breadth | Newest methods, full control |
| RL methods | DPO, KTO, ORPO, GRPO | DPO, GRPO | PPO, DPO, KTO, ORPO, SimPO | GRPO, RLOO, DPO, KTO |
| License | Apache-2.0 | Apache-2.0 core, AGPL-3.0 UI | Apache-2.0 | Apache-2.0 |
Sources: GitHub · Docs · axolotl.ai · Ultradune comparison (2026)
Is this your tool? Add a badge