Which models and tuning methods does LLaMA Factory support?

It supports over 100 open models, including LLaMA, Mistral, Qwen, and multimodal VLMs. Tuning methods include full-parameter, LoRA variants, QLoRA, DPO, PPO, KTO, and ORPO.

Is LLaMA Factory free to use?

Yes, it is free and open-source under the Apache-2.0 license. It is self-hosted with no SaaS tier, meaning you must provide your own GPU hardware.

What acceleration and export options are available?

The tool offers acceleration via FlashAttention-2, Unsloth, Liger Kernel, and vLLM for inference. It also supports exporting models to GGUF and Ollama for downstream serving.

Who is LLaMA Factory designed for?

It is aimed at practitioners and small labs who want to iterate on local GPUs or rented hardware without manually configuring TRL, PEFT, or DeepSpeed.

What are the main limitations of using LLaMA Factory?

Users must bring their own GPUs and handle operations. The WebUI abstracts but does not eliminate VRAM-planning challenges, and rapid releases require careful version pinning.

LLaMA Factory

Open-source, no-code WebUI for fine-tuning 100+ open LLMs with LoRA, QLoRA, DPO, and PPO.

Free· Free, open-source (Apache-2.0); self-hostedFine-tuningMulti-model (LLaMA, Mistral, Qwen, Gemma, Phi, LLaVA, ChatGLM, Yi)7.2 / 10

Visit website →

In short

LLaMA Factory is an open-source platform that simplifies post-training for open-weight LLMs via a no-code WebUI. It is best for practitioners who want to fine-tune models like LLaMA or Mistral on their own hardware without writing custom scripts.

Best for

Pick LLaMA Factory if you want one tool to fine-tune any open-weight LLM on your own hardware without writing custom training scripts.

Skip if

Skip it if you want a managed, click-to-train cloud service or don't have access to suitable GPUs.

LLaMA Factory is an open-source training and fine-tuning platform that wraps the messy reality of post-training open-weight LLMs behind a clean Python toolkit and a LlamaBoard WebUI. It supports a wide model zoo (LLaMA, Mistral, Qwen, Yi, Gemma, Baichuan, ChatGLM, Phi, LLaVA and other VLMs) and covers the full pipeline: continued pre-training, supervised instruction tuning, reward modeling, and preference alignment via PPO, DPO, KTO, and ORPO.

What sets it apart is the breadth of tuning methods exposed without code: full-parameter, freeze-tuning, LoRA and its variants (LoRA+, DoRA, PiSSA, rsLoRA), plus 2/3/4/5/6/8-bit QLoRA via bitsandbytes, HQQ, or AQLM. Acceleration paths include FlashAttention-2, Unsloth, Liger Kernel, and vLLM for inference. It is aimed at practitioners and small labs who want to iterate on local GPUs (or a single rented H100) instead of stitching together TRL, PEFT, accelerate, and DeepSpeed configs by hand. It is fully free and self-hosted, with no SaaS tier.

Integrations cover Hugging Face Hub for models and datasets, TensorBoard / Weights & Biases / MLflow / SwanLab for experiment tracking, and exports to GGUF and Ollama for downstream serving. Caveats: you still need your own GPU(s), the WebUI hides but does not eliminate VRAM-planning headaches, and the project moves fast enough that pinned versions matter.

Editor's take

This is the de-facto Swiss army knife for open-LLM fine-tuning in 2025-2026 — broader model coverage than Axolotl, friendlier than raw TRL/PEFT, and one of the few projects that keeps pace with new architectures within days. Best paired with Unsloth for single-GPU runs and vLLM for serving.

— The AI Tool Bible editorial team

Pros

✅ No-code WebUI (LlamaBoard) covers SFT, DPO, PPO, KTO, and reward modeling
✅ Supports 100+ open models including multimodal VLMs out of the box
✅ Full QLoRA stack (2-8 bit) plus LoRA+, DoRA, PiSSA variants
✅ Acceleration via FlashAttention-2, Unsloth, Liger Kernel, vLLM inference
✅ Exports to GGUF / Ollama and integrates with W&B, MLflow, TensorBoard

Cons

⚠️ Self-hosted only — you bring the GPUs and the ops
⚠️ Rapid release cadence means version pinning is essential
⚠️ WebUI abstracts but does not solve VRAM and dataset-formatting pitfalls

Use cases

lora-fine-tuningqloradpo-alignmentinstruction-tuningrlhfvlm-fine-tuning

Frequently asked

Which models and tuning methods does LLaMA Factory support?: It supports over 100 open models, including LLaMA, Mistral, Qwen, and multimodal VLMs. Tuning methods include full-parameter, LoRA variants, QLoRA, DPO, PPO, KTO, and ORPO.
Is LLaMA Factory free to use?: Yes, it is free and open-source under the Apache-2.0 license. It is self-hosted with no SaaS tier, meaning you must provide your own GPU hardware.
What acceleration and export options are available?: The tool offers acceleration via FlashAttention-2, Unsloth, Liger Kernel, and vLLM for inference. It also supports exporting models to GGUF and Ollama for downstream serving.
Who is LLaMA Factory designed for?: It is aimed at practitioners and small labs who want to iterate on local GPUs or rented hardware without manually configuring TRL, PEFT, or DeepSpeed.
What are the main limitations of using LLaMA Factory?: Users must bring their own GPUs and handle operations. The WebUI abstracts but does not eliminate VRAM-planning challenges, and rapid releases require careful version pinning.