The ModelScope team at Alibaba develops ms-swift, the Scalable lightWeight Infrastructure for Fine-Tuning.

A framework for pretraining, supervised finetuning, RLHF alignment, inference, evaluation, and deployment of 600+ LLMs and 300+ multimodal models.

Runs on user-managed GPU infrastructure as a Python package, integrated with the ModelScope ecosystem.

Actively developed since 2023, with continuous updates supporting newer model families through 2026.

Gives teams a single toolkit spanning LoRA, QLoRA, full-parameter tuning, DPO, and GRPO across text and multimodal models.

Built on PyTorch and Hugging Face-compatible tooling, and integrates with ModelScope's model hub.

Open-source (Apache-2.0 license) on GitHub, self-hosted training and inference, no external data transmission required.