Model components / Fine-tuning code

SWIFT

Alibaba ModelScope's unified PEFT/full-parameter CPT/SFT/DPO/GRPO toolkit covering 600+ LLMs and 300+ MLLMs (Qwen3, DeepSeek-R1, GLM, Llama4, InternVL, etc.). Picked over LLaMA-Factory when working in the ModelScope ecosystem or when broad multimodal model coverage matters. SWIFT often ships day-zero recipes for Chinese-origin models. 14.2K GitHub stars, 147 contributors, 335 commits in 90 days; published as AAAI 2025.

ms-swift (SWIFT) by ModelScope/Alibaba, v4.2.3 (patch) released May 31 2026. Fine-tuning + deployment framework covering 600+ LLMs and 300+ multimodal models. Confirmed live on GitHub June 2026.

Openness

5 high confidence

5.0

license: Apache-2.0(OSI)
source: public
maintainer: ModelScope(Alibaba)
core-gated: ungated

Fully OSI-licensed open source training/deploy framework; source public.

https://github.com/modelscope/ms-swift recorded 2026-06-04
Apache-2.0 license; v4.2.3 (May 31 2026); ModelScope/Alibaba maintainer

Adoption

4 high confidence

4.0

~110k PyPI downloads last month (primary). Broad coverage (600+ text + 300+ multimodal models incl. Qwen3, DeepSeek, Llama, GLM) and strong uptake in the Chinese/ModelScope ecosystem. Lower bound of the 100K-1M band; level 4 on sustained monthly download volume.

https://pypistats.org/packages/ms-swift recorded 2026-06-04
~110,083 PyPI downloads last month
https://github.com/modelscope/ms-swift recorded 2026-06-04
600+ text models, 300+ multimodal models supported

Capability

4 high confidence

4.0

Very broad method + model coverage and an end-to-end train/eval/deploy pipeline. Scored 4 (not 5) vs verl/Megatron: it is a comprehensive practitioner toolkit rather than the demonstrated large-scale frontier-training definer.

https://github.com/modelscope/ms-swift recorded 2026-06-04
LoRA/QLoRA/DoRA, SFT, GRPO family, DPO/KTO/CPO/SimPO/ORPO, pretraining across 600+/300+ models

Unchanged since 2026-07-30 (last edited, not re-checked)