Axolotl

What it is for

Axolotl is an Apache-2.0 post-training toolkit: one YAML drives preprocess, train, eval, and export for SFT, preference, or RL. Use it when the job is fine-tune an open model on a GPU you already have, not when you need a hosted chat seat.

What makes it unique: the cover is the official axolotl-ai-cloud/axolotl GitHub file list — src/axolotl/, examples/, docs/, Apache-2.0 in the About pane — not a wordmark.

Key features

• YAML SFT train — axolotl fetch examples then axolotl train examples/llama-3/lora-1b.yml (Quickstart). LoRA / QLoRA / full / QAT are adapters on this job, not five products. Multipacking, Flash Attention, FSDP2, and DeepSpeed sit on the same train.
• Preference tuning — DPO, IPO, KTO, ORPO, SimPO on chosen/rejected (or KTO binary) pairs (rl: dpo / orpo / kto; Support matrix). A different data shape, not SFT with a flag.
• GRPO / online RL — generate-then-reward loops (rl: grpo, GDPO) with a reward function (choosing a method). An online rollout, not a static pair.
• Reward models — outcome RM or stepwise PRM so something else can score a reply (reward_model / process_reward_model; Support matrix). A scorer, not the policy train.
• Export — axolotl export (GGUF, optional quant) and axolotl merge-lora so you can hand weights to Ollama or llama.cpp (docs home; Quickstart merge). A package, not a train.

Multimodal VLM / audio datasets are input variants on the same train. uv pip install axolotl / Docker are install / SDK. You bring the GPU bill — that is the meter. Base-model names are SKU.

Who it's for

Best fit: ML engineers who already have a GPU and will write a YAML + a JSONL, not a team looking for a $20 chat plan.

Pricing

Verified 2026-10-22 (Asia/Shanghai) from the Apache-2.0 LICENSE and docs.axolotl.ai. There is no hosted SKU table.

• Software — Apache 2.0. License fee $0.
• GPU / cloud bill — you pay RunPod / Vast / the electricity. Axolotl does not print those dollars.

Do not invent an Axolotl Cloud seat. Dedicated support is a quote (wing@axolotl.ai), not a public card.

Limitations / costs

• You bring the GPU. The trainer is $0; A100 hours are not.
• VRAM is the meter. A 70B QLoRA still needs a large card.
• Fine-tune CLI, not a Studio chat. For a local Studio form, see Unsloth.

Alternatives

• Unsloth — faster LoRA kernels / Studio when you already live in notebooks.
• LLaMA-Factory — a Web UI trainer when YAML is not the desk you want.
• Hugging Face — Hub + TRL / AutoTrain when you want their hosted path.

Official tutorial

We searched for a first-party Axolotl YouTube getting-started lesson we could oEmbed. Community walkthroughs (Fahd Mirza, BittyBattyBunny) are not embedded. Use written Quickstart (axolotl fetch examples then axolotl train examples/llama-3/lora-1b.yml).

Website

https://axolotl.ai

Repository

https://github.com/axolotl-ai-cloud/axolotl

Source

Catalog row curated for AI Tools Nav. Cover is the official axolotl-ai-cloud/axolotl GitHub file list (Apache-2.0 + src/axolotl). License fee $0; no hosted seat table. No first-party YouTube embed.

Last verified

2026-10-22 (Asia/Shanghai) editorial check.