LlamaFactory iconLlamaFactory

Open-source toolkit to fine-tune 100+ LLMs and vision-language models from a web UI or CLI, with LoRA, QLoRA, DPO, PPO and vLLM serving

76k+ GitHub Stars
100+ LLMs and VLMs
ACL 2024 Peer-Reviewed Paper

Overview

LlamaFactory (formerly LLaMA-Factory; the GitHub repo was renamed) offers "unified efficient fine-tuning of 100+ LLMs and VLMs." Its standout is LLaMA Board, a Gradio web UI where you pick a model, dataset and method and start training without writing code; the same jobs also run from a CLI. It supports full, freeze, LoRA, QLoRA and OFT tuning, and training stages from pre-training and SFT to reward modeling, PPO, DPO, KTO, ORPO and SimPO. Trained models can be served through vLLM or SGLang behind an OpenAI-style API. The project is led by Yaowei Zheng and described in a paper accepted to the ACL 2024 demo track; its README lists Amazon, NVIDIA and Aliyun among users. Code is Apache-2.0. The latest release, v0.9.5 (May 30, 2026), added Qwen3.5/3.6, Gemma 4 and Transformers v5 support.

The Verdict

Who Should Use LlamaFactory?

Best For

  • Newcomers who want to fine-tune from a browser UI
  • Teams trying many model families and methods quickly
  • Researchers who want one toolkit from SFT to PPO and DPO

Not Ideal For

  • Large multi-node production runs (Axolotl is stronger there)
  • Teams that need frequent releases: the last tag is from May 2026
  • Buyers who need vendor support

What's Great

  • No-code web UI plus CLI for the same jobs
  • Wide model and method coverage, peer-reviewed design
  • Built-in vLLM and SGLang inference with an OpenAI-style API
  • Very large community (see badge)

Watch Out For

  • README warns that look-alike third-party sites are not official
  • Parts of the docs are in Chinese only
  • Slower release pace: no tagged release since May 30, 2026

Pricing

View all features & details

Key Features

  • Tuning modes: full, freeze, LoRA, QLoRA, OFT, QOFT
  • Stages: pre-training, SFT, reward modeling, PPO, DPO, KTO, ORPO, SimPO
  • LLaMA Board web UI (Gradio)
  • vLLM and SGLang inference, OpenAI-style API server
  • Model families incl. Llama, Qwen, Gemma, DeepSeek, Mistral, Phi, LLaVA

Project

  • Lead author Yaowei Zheng (GitHub: hiyouga)
  • Paper: arXiv 2403.13372, ACL 2024 System Demonstrations
  • v0.9.4 Dec 31, 2025; v0.9.5 May 30, 2026

Real-World Usage

Named Users

  • Amazon, NVIDIA and Aliyun (README)
  • Dozens of listed projects built with it, such as Sky-T1

Research

  • Peer-reviewed at ACL 2024
  • Seven authors led by Yaowei Zheng

How It Compares

Feature LlamaFactory Axolotl Unsloth
Interface Web UI + CLI YAML config + CLI Python, Studio web UI
Strongest at Easy start, model breadth Multi-GPU, multi-node Single-GPU speed and memory
RL methods PPO, DPO, KTO, ORPO, SimPO DPO, KTO, ORPO, GRPO DPO, GRPO
Built-in serving vLLM, SGLang Inference from the same YAML Run models in Unsloth Studio or Desktop

Sources: GitHub · Releases · Docs · Paper

Is this your tool? Add a badge