Verified project record

modelscope/ms-swift

ms-swift is a unified framework for fine-tuning, deploying, evaluating, and quantizing text and multimodal large language models. It supports 600+ text-only models and 400+ multimodal models across training, inference, evaluation, quantization, and deployment stages.

Project overview

ms-swift consolidates the entire model lifecycle pipeline from training through deployment into a single framework, covering 600+ text-only and 400+ multimodal models with both lightweight and distributed training support.

Project type: Model Development
Deployment: Refer to project documentation
License: Apache-2.0

Best for

AI engineers who need a single framework for training, evaluating, quantizing, and deploying large language models
Researchers working with both text-only and multimodal models who want lightweight or distributed training options
Developers who prefer CLI, library, or Gradio-based Web UI interaction for model lifecycle tasks

Key capabilities

Supports pre-training, full-parameter fine-tuning, and lightweight training (LoRA, QLoRA, DoRA, etc.) for text and multimodal models.
Performs inference using Transformers, vLLM, SGLang, or LMDeploy engines, including interactive command-line and interface inference.
Provides a Gradio-based Web-UI interface for training, inference, evaluation, and quantization.

Limitations and risks

A GPU is not strictly required but is needed for optimal performance; minimum documented hardware is a single 3090 for 4B fine-tuning or 9GB training resources for 7B quantized training.
Model and dataset management depends on ModelScope or HuggingFace Hub; network access to these services is needed for downloading and pushing models and datasets.
Setup requires a Python and PyTorch environment, placing the difficulty at medium for initial configuration.

Getting started

Install with pip (pip install ms-swift -U), then run a first supervised fine-tuning command such as: swift sft --model Qwen/Qwen3-4B-Instruct-2507 --dataset AI-ModelScope/alpaca-gpt4-data-zh. Setup requires a Python and PyTorch environment, and a GPU is recommended for optimal performance.
A single 3090 GPU suffices for 4B model fine-tuning, and 9GB of training resources can support 7B models using quantized training.

Evidence and sources

Open original GitHub project