An open source fine-tuning (post-training) tool developed by Axolotl AI. It runs on your own GPU machine or on a cloud GPU instance, so you never have to hand your models or training data over to a third-party service. Its defining feature is that a training run is configured in a single YAML file: write down the model, dataset, training method, batch size and so on, run one command, and training starts. Since its release in May 2023 it has become one of the standard tools in the LLM fine-tuning community, and it is licensed under Apache 2.0.
Key Features
- Training driven entirely by a YAML config: Put the model name, dataset, training method and hyperparameters in a YAML file and run something like
axolotl train config.yml. There is no need to write your own training script, and you can compare setups simply by swapping config files - Broad support for major open models: Works with major models on the Hugging Face Hub such as LLaMA, Mistral, Mixtral, Qwen, Gemma, Phi and GPT-OSS. Beyond text, it also supports training vision-language models and audio models
- A wide range of training methods: In addition to full fine-tuning, it covers LoRA and QLoRA for training with limited GPU memory, preference tuning methods such as DPO, IPO, KTO and ORPO, reinforcement learning with GRPO, and reward modeling
- Multi-GPU and multi-node distributed training: Supports multi-GPU training with FSDP, DeepSpeed and DDP, and multi-machine training via Torchrun and Ray. Distributed techniques aimed at larger models, such as sequence parallelism and expert parallelism, are available as well
- Your data stays with you: It runs on your own servers or in the cloud, under Docker or Kubernetes. Training data and the resulting model weights remain in your hands, which makes it easier to work with confidential data
- Many dataset formats: Handles pre-training, instruction tuning, conversation, template-free and pre-tokenized formats, loading from local storage, Hugging Face, or cloud storage such as S3
Pricing
| Plan | Price | Highlights |
|---|---|---|
| Open source | $0 (Apache 2.0) | All features available at no cost, with no functional limits |
| Dedicated support | Contact for pricing | Onboarding and operational support from the developers (via the contact channel on the official site) |
Pricing is as of August 2026. Please check the official site for the latest information. Axolotl itself is free, but the GPUs used for training — your own machines or cloud GPU instances — are a separate cost.
Pros & Cons
✅ Pros
- The software itself is completely free with no functional limits
- A single YAML file holds the whole setup, making training conditions easy to record, share and reproduce
- Wide coverage of models and training methods, from LoRA through to reinforcement learning, under one framework
- Models and data stay local, which suits sensitive datasets
- Scales from a single GPU to multiple nodes as an extension of the same config file
⚠️ Cons
- No GUI; it assumes you are comfortable with the command line and editing YAML, which raises the bar for complete beginners
- You must supply the GPUs yourself, and costs can be significant depending on model size
- Requires an NVIDIA GPU (Ampere generation or newer) or an AMD GPU, along with specific Python and PyTorch version requirements
- Development moves quickly and config options can change, so pinning a version is the safer approach
- Results depend heavily on data quality; installing the tool alone does not guarantee a good model
Comparison with Similar Services
| Criteria | Axolotl | Unsloth | LLaMA-Factory | TRL (Hugging Face) |
|---|---|---|---|---|
| Provider | Axolotl AI | Unsloth AI | Individual and community driven | Hugging Face |
| Interface | YAML config + CLI | Python code / notebooks | GUI + CLI | Python library |
| Strength | Post-training in general, including preference tuning and reward modeling | Fast training with limited GPU memory | Ease of use, GUI-based operation | Integration with the Hugging Face ecosystem |
| Distributed training | Broad support: FSDP, DeepSpeed, Ray and more | Mainly single GPU | Supported | Supported |
| Price | Free (OSS) | Free (OSS, paid tier available) | Free (OSS) | Free (OSS) |
As of 2026, the core capabilities — LoRA, QLoRA, DPO, GRPO and vision-language models — are present in all of these tools. The differences lie in workflow: how you configure a run, how you distribute it, and how you reproduce it.
Who Is It For
- Developers who want to build an in-house LLM on their own data without writing a training script from scratch
- Companies and research teams that cannot send confidential data to an external fine-tuning service
- Anyone who wants to go beyond LoRA and experiment with preference tuning and reinforcement learning such as DPO and GRPO
- Teams that expect to move to multi-GPU or multi-node training at some point
- People who want training conditions saved as files so experiments stay reproducible
Summary
Axolotl is an open source tool that condenses the work of LLM fine-tuning into a single YAML config file. Its strengths are broad coverage — from full fine-tuning and LoRA/QLoRA to preference tuning and reinforcement learning — and the ability to move from one GPU to multi-node distributed training within the same framework. There is no GUI and it assumes command line use, but if you want to train while keeping your models and data local, it is a strong option. The easiest way in is to run one of the small-model config examples from the Getting Started section of the official docs, then swap in your own data.