MS-Swift is the ModelScope community’s framework for fine-tuning and deploying large language models and multimodal models at scale. It supports training (pre-training, supervised fine-tuning and human alignment methods such as DPO and GRPO), inference, evaluation, quantization and deployment for more than 600 text LLMs (Qwen, DeepSeek, GLM, Llama, InternLM and many more) and hundreds of multimodal models, using either parameter-efficient methods like LoRA or full-parameter training. The framework was presented at AAAI 2025 and has gathered nearly 16,000 stars on GitHub, with active development in 2026.
Fine-tuning a modern LLM normally means gluing together model-specific scripts, training loops and serving tools. MS-Swift standardizes the whole journey in one framework and one command-line interface: point it at a model and a dataset, choose PEFT or full-parameter training, and it handles distributed training (including Megatron parallelism), alignment algorithms, evaluation benchmarks and export for deployment with engines such as vLLM. Because it tracks the fast-moving open-model ecosystem, new releases like the latest Qwen, DeepSeek and GLM generations are typically supported quickly, which has made it a default choice for teams fine-tuning open models, especially in the ModelScope ecosystem.
MS-Swift is free and open source under the Apache-2.0 license. Your only costs are the compute you train on and the models you choose to fine-tune.
MS-Swift is for ML engineers and researchers who fine-tune open-weight models and want one maintained framework covering training, alignment, evaluation and deployment instead of a pile of per-model scripts.