Skip to content
ReviewRamble

Axolotl

Config-driven fine-tuning tool for open LLMs, from single GPUs to multi-node clusters.

Reviewed by Sheela · Updated

View on GitHub Open source

Our rating

4.3/5

Axolotl is where fine-tuning goes once one GPU stops being enough: one YAML file, every parallelism strategy, and a learning curve to match.

What it is

Axolotl is an open-source post-training framework for LLMs driven by a single YAML file: you set the base model, dataset format and method, then run `axolotl train`. The same config covers preprocessing, training, evaluation, quantisation and inference, and `axolotl export` now produces GGUF files for llama.cpp and Ollama.

It covers full fine-tuning, LoRA and QLoRA, quantisation-aware training down to NVFP4 and MXFP4, preference tuning (DPO, IPO, KTO, ORPO), reinforcement learning (GRPO, GDPO) and reward models, for text, vision-language and audio models. Scaling uses DeepSpeed, FSDP2 and composable tensor, context and expert parallelism across nodes.

It needs an Nvidia GPU (Ampere or newer for bf16 and Flash Attention) or an AMD GPU, Python 3.12 and PyTorch 2.13 or later. Version 0.21.0 was released on 16 October 2026 under Apache 2.0.

Our review

Axolotl's case rests on reproducibility at scale. A run is one YAML file that someone else can rerun, and the same format stretches from a LoRA on a single card to multi-node training that composes FSDP2 with tensor, context and expert parallelism. At about 12,600 GitHub stars it is far smaller than Unsloth (about 78,500) or LlamaFactory (about 75,700), but its configs turn up on published model cards because they are easy to share.

Speed of model support is its other selling point. In September and October 2026 alone the README lists MiniCPM5, Qwen3.6 and Reflection's Beam, CPU-staged FSDP2 QLoRA loading for models larger than one GPU's memory, GGUF export, and an expert-parallel mode that trains mixture-of-experts models over NVLink or PCIe without DeepEP. Async GRPO, added in April, is documented as up to 58% faster per step.

Telemetry is on by default, which matters on shared clusters. Axolotl sends PostHog its system and hardware details, runtime metrics, whitelisted model names and configurations, and sanitised error traces, and pauses start-up for 10 seconds to say so unless you set AXOLOTL_DO_NOT_TRACK=1. The docs say no personal data or file paths are collected.

The cost is expertise. A long config reference, version pins on Python and PyTorch, and distributed-training vocabulary are assumed knowledge, and there is no web interface of the kind LlamaFactory offers; the Docker image removes most environment trouble. For a first LoRA on one card, Unsloth and LlamaFactory are gentler starts. Past that point Axolotl's range of parallelism is the reason to accept the curve.

Where it earns its place

  • An ML engineer at a legal-tech start-up renting an 8-GPU node

    Fine-tune a large open model on contract data with reproducible runs

    Write one YAML with QLoRA and FSDP2, use CPU-staged loading so each GPU need not hold a full quantised copy, and commit the file beside the dataset version. Axolotl is free under Apache 2.0; the bill is GPU hours, and the config makes reruns after a data refresh routine.

  • A researcher adding reasoning behaviour to a small model

    Run GRPO with a custom reward function on a workstation GPU

    Start from the bundled examples (axolotl fetch examples), switch the method to GRPO and point it at the reward function; async GRPO is documented as up to 58% faster per step. It needs an Ampere-or-newer Nvidia card for bf16 and Flash Attention.

  • A product team at a 30-person company shipping an on-device assistant

    Turn a fine-tuned model into something that runs in Ollama

    Train a LoRA, merge it, then use axolotl export to write a quantised GGUF for llama.cpp or Ollama, replacing a separate conversion toolchain. Set AXOLOTL_DO_NOT_TRACK=1 if configurations or model choices are commercially sensitive.

Before you commit

Setup
uv-based install (the matching PyTorch build, then `uv pip install --no-build-isolation axolotl[deepspeed]`) or the axolotlai/axolotl Docker image, which the README calls less error-prone. `axolotl fetch examples` and `axolotl train examples/llama-3/lora-1b.yml` give a first run; a Colab notebook covers the no-hardware case.
Integrations
Hugging Face models and datasets; S3, Azure, GCP and OCI data; DeepSpeed and FSDP2; GGUF export for llama.cpp and Ollama; cloud launch plugins including Nebius Serverless Jobs and templates on RunPod, Vast.ai and Modal; bundled agent docs via `axolotl agent-docs`.
Data and privacy
Runs on your hardware. Opt-out PostHog telemetry sends system, hardware, runtime metrics, whitelisted model names, configurations and sanitised errors; AXOLOTL_DO_NOT_TRACK=1 disables it. Apache 2.0 allows commercial use, though base-model licences still apply. No security advisories are published on the repo.
Cost at scale
The software is free; the cost is GPUs. Axolotl publishes no per-model VRAM table, so size jobs from a trial run; multi-node work bills at your cloud's GPU-hour rate. Dedicated support is available by arrangement with the maintainers, at prices not published.
Limits
No GUI; Python 3.12 and PyTorch 2.13+ requirements; about 250 open issues; multimodal training launched as beta.

What's good

  • One reproducible YAML file per run
  • FSDP2, DeepSpeed and N-D parallelism across nodes
  • SFT, DPO, GRPO, QAT and reward models
  • Quick support for new models; GGUF export

What's not

  • No graphical interface
  • Telemetry on by default
  • Assumes distributed-training knowledge
  • Nvidia or AMD GPU only

Best for

Practitioners running repeatable fine-tunes, especially across several GPUs.

Alternatives

Unsloth

Training & fine-tuningPython78k stars

4.6/5

Faster, lower-memory fine-tuning for open LLMs, now with a desktop app and web UI for running and training models.

TRL

Training & fine-tuningPython20k stars

4.5/5

Hugging Face's library for post-training LLMs: supervised fine-tuning, DPO, GRPO and other reinforcement learning methods.

LlamaFactory

Training & fine-tuningPython76k stars

4.3/5

Fine-tuning framework for 100+ LLMs and vision-language models, with a web UI as well as a command line.

Axolotl: common questions

Is Axolotl worth it?
Axolotl is where fine-tuning goes once one GPU stops being enough: one YAML file, every parallelism strategy, and a learning curve to match. We rate it 4.3 out of 5.
How much does Axolotl cost?
Free and open source under the Apache 2.0 license.
What are the downsides of Axolotl?
No graphical interface; telemetry on by default; assumes distributed-training knowledge; nvidia or AMD GPU only.
Who is Axolotl best for?
Practitioners running repeatable fine-tunes, especially across several GPUs.
How much does Axolotl cost for a team or in production?
The software is free; the cost is GPUs. Axolotl publishes no per-model VRAM table, so size jobs from a trial run; multi-node work bills at your cloud's GPU-hour rate. Dedicated support is available by arrangement with the maintainers, at prices not published.
Is Axolotl safe for company data?
Runs on your hardware. Opt-out PostHog telemetry sends system, hardware, runtime metrics, whitelisted model names, configurations and sanitised errors; AXOLOTL_DO_NOT_TRACK=1 disables it. Apache 2.0 allows commercial use, though base-model licences still apply. No security advisories are published on the repo.
What are the best Axolotl alternatives?
The closest alternatives we have reviewed are Unsloth, TRL and LlamaFactory.

Answers checked Oct 23, 2026.

Make Axolotl? Show this rating on your siteAxolotl rated 4.3 out of 5 on ReviewRamble