Axolotl
Axolotl AIAxolotl is a YAML-config-driven fine-tuning framework built on Hugging Face's TRL and PEFT, plus bitsandbytes for quantization, that runs full fine-tuning, LoRA, QLoRA, DPO, and GRPO from a single configuration file rather than a custom training script. It scales from a single GPU to multi-node clusters through FSDP, DeepSpeed, Torchrun, and Ray, and supports models from across the Hugging Face Hub. Axolotl AI builds and maintains it.
Openness
5 high confidence- license
- Apache-2.0(OSI)
- source
- public(axolotl-ai-cloud/axolotl)
- core features withheld
- no — Apache-2.0 across the repo, no ee/ or enterprise directory, no license key and no closed package
Apache-2.0 across the whole repository, with the full source public and no proprietary managed training service, only third-party cloud deployment templates. Nothing is withheld from the open package: no enterprise directory, no license-key gate, and the only paid offering is dedicated support by email.
- https://axolotl.ai/ recorded 2026-09-28
The homepage describes Axolotl as open source ("Axolotl is an Open Source tool making fine-tuning AI models faster") that runs "in your own cloud, with your own Docker and k8s setup". Searched the body for pricing and price: none found.
- https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/LICENSE recorded 2026-09-28
The file is the standard Apache License text: "Apache License" and "Version 2.0, January 2004", with the grant of a "perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable" license.
- https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/README.md recorded 2026-09-28
The README, served from the public repository, states "Axolotl is a free and open-source tool designed to streamline post-training and fine-tuning" and "This project is licensed under the Apache 2.0 License". It documents no paid tier or license key; searched the body for enterprise, pricing, license key and hosted tier: none. The only commercial line is "Need dedicated support? Please contact".
Adoption
2 medium confidenceAxolotl is a popular config-driven fine-tuning tool in the open-model community, and PyPI downloads in the trailing month are the measure used. That volume is modest next to TRL, PEFT and Unsloth, which keeps it below their reach; GitHub stars, tracked separately, corroborate but do not set the figure.
- https://github.com/axolotl-ai-cloud/axolotl recorded 2026-08-09
stargazerCount 12330 (corroborating signal, not the adoption basis)
- https://pypi.org/pypi/axolotl/json recorded 2026-08-09
info.name = axolotl, info.version = 0.18.0: a live PyPI project named axolotl exists, even though the product file's artifact list carries no pypi entry and the warehouse signal_pypi table has no row for it (see findings).
- https://pypistats.org/packages/axolotl recorded 2026-08-09
Downloads last month: 12,389
Capability
4 high confidenceAxolotl runs full fine-tuning, LoRA, QLoRA and quantization-aware training, preference methods such as DPO, KTO and ORPO, GRPO reinforcement learning and reward modeling from one configuration file, with FSDP2, DeepSpeed, sequence and expert parallelism and multi-node training through Torchrun or Ray. It trails Megatron-LM on demonstrated scale, publishing no training run near the size of Megatron-LM's multi-thousand-GPU benchmarks.
- https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/README.md recorded 2026-09-28
The README lists the training methods "Full fine-tuning, LoRA, QLoRA, GPTQ, QAT (int8/int4/FP8/NVFP4/MXFP4)" and "Preference Tuning (DPO, IPO, KTO, ORPO), RL (GRPO, GDPO), and Reward Modelling (RM) / Process Reward Modelling (PRM)", plus "Multi-GPU training (FSDP2, DeepSpeed)" and "Multi-node training (Torchrun, Ray)", Sequence Parallelism, ND Parallelism and Expert Parallelism.
Verified 2026-09-28, except some scores, still awaiting re-confirmation