Dev Tools
peft
Fine-tune huge models on small hardware with LoRA and other efficient methods
★ 21.8k starsApache-2.0
View guide →Fine-tune and deploy 600+ LLMs and 300+ multimodal models with PEFT or full-parameter training.
ms-swift is an open-source dev tools project by modelscope (modelscope/ms-swift on GitHub). It has 15,796 GitHub stars and 1,726 forks, is released under the Apache-2.0 license, and was last updated on GitHub on 8 October 2026.
Choose from 600+ LLMs or 300+ multimodal models, then decide between lightweight PEFT (such as LoRA) and full-parameter training.
Install the package from PyPI with pip or uv, or clone the repo and install from source for the latest main branch.
Use one of the 150+ built-in datasets or format your own for pre-training, SFT, alignment, embedding or reranker tasks.
Launch training with the CLI or scripts, evaluate the result, then quantize and serve it with vLLM, SGLang or LMDeploy.
Training-to-deployment pipelines for 600+ text LLMs and hundreds of multimodal models, with day-0 support for popular new releases.
CPT, SFT, DPO, KTO, RM, CPO, SimPO and ORPO, plus a GRPO-family reinforcement learning set including DAPO, GSPO and Reinforce++.
LoRA, QLoRA, DoRA, LoRA+, LLaMAPro, LongLoRA, Adapter, ReFT and more, alongside full-parameter training.
Tensor, pipeline, context and expert parallelism techniques from Megatron to accelerate large training runs.
NVIDIA GPUs via CUDA, AMD GPUs via ROCm, Ascend NPU and other domestic accelerators, plus Apple MPS and CPU.
Train on BNB, AWQ, GPTQ, AQLM, HQQ and EETQ quantized models, and accelerate inference and deployment with vLLM, SGLang and LMDeploy.
Clone ms-swift if installing from source; PyPI installs do not need the clone.
git clone https://github.com/modelscope/ms-swift.git
cd ms-swiftInstall the latest release from PyPI and upgrade to the newest version.
pip install ms-swift -UFor the main branch, install the cloned project in editable mode.
pip install -e .It is provided by the ModelScope community (modelscope) and open-source contributors, and was presented at AAAI 2025.
Yes. It is Apache-2.0 licensed; the models and datasets you use with it carry their own licenses.
No. Training and inference run on your own hardware; you only need model weights and data, which you download from hubs such as ModelScope or Hugging Face.
All three fine-tune open LLMs; ms-swift stands out for its very wide model and multimodal coverage, Megatron parallelism and the breadth of built-in alignment algorithms - the best fit depends on your model and hardware.
Run pip install ms-swift -U, pick a built-in dataset and a supported base model, and follow the quick-start training example in the docs.
Fine-tune huge models on small hardware with LoRA and other efficient methods
20+ high-performance LLMs implemented from scratch, with proven recipes to pretrain, finetune and deploy.
A streamlined framework for evaluating and benchmarking LLMs, VLMs and AIGC models.
Resource hub for AI-powered game development — LLMs, world models, agents.
Unified, efficient fine-tuning of 100+ LLMs and VLMs with a CLI and Web UI — ACL 2024
AI enabled pair programmer for Claude, GPT, O Series, Grok, Deepseek, Gemini and 300+ models
The most-starred other Dev Tools projects in the LibreAI catalogue — the realistic shortlist if ms-swift is not the right fit.
The open source coding agent
The model-definition framework behind modern AI — thousands of pretrained transformers for text, vision, and audio.
Anthropic's agentic coding tool that lives in your terminal and works across your whole codebase
Cross-platform desktop all-in-one manager for Claude Code, Codex, OpenCode and other coding agents
OpenAI's lightweight coding agent that runs locally in your terminal
Free and open — grab the source, star the repo, and start creating.