The latest release of modelscope/ms-swift is v4.6.0. It is a training and deployment framework for large language models (LLMs) and multimodal models provided by the ModelScope community, written in Python and released under the Apache-2.0 license. However, the individual changes in v4.6.0 are not broken down in this record.
Its broad coverage stands out first. Continued pre-training (CPT), supervised fine-tuning (SFT), alignment training (DPO), GRPO-family reinforcement learning algorithms, inference, evaluation, quantization and deployment are all handled by a single CLI. The README states support for more than 600 LLMs and more than 300 multimodal models, and cites language models such as Qwen3.8, DeepSeek-V4 and GLM-5.1 alongside multimodal models such as Qwen3-VL and InternVL3.5. It offers both full-parameter training and lightweight fine-tuning such as PEFT and LoRA, and a web user interface can run training, inference, evaluation and quantization. The repository description lists AAAI 2025.
However, the evaluation records contain no performance or quality scores, only a momentum metric of 4.4 (out of 10). No own benchmarks showing training speed or stability are presented either, so judging by the numbers how well it works is premature. It is worth a first look for those bundling several models into a single CLI pipeline; for now, the only certain thing is the feature scope stated in the README.