SWIFT

license

Apache-2.0(OSI)

source

public

maintainer

ModelScope(Alibaba)

core-gated

ungated

Apache-2.0, an OSI license, with the full source of the training and deployment framework public and nothing held back for a paid tier.

• https://github.com/modelscope/ms-swift recorded 2026-08-08

  Repository tree lists swift/, setup.py, requirements/, tests/, docs/ and LICENSE, and the README install section gives 'git clone https://github.com/modelscope/ms-swift.git / cd ms-swift / pip install -e .', so the running framework builds from the published repo. The README License section reads 'This framework is licensed under the Apache License (Version 2.0)' and the sidebar Resources block lists 'Apache-2.0 license'. The README documents the complete pipeline (swift pt / sft / rlhf / infer / app / deploy / sample / eval / export, plus 'swift web-ui') with no paid, premium, enterprise or commercial edition; the only 'Enterprise', 'Premium' and 'Pricing' strings on the page sit in GitHub's own site navigation above the repository header, not in project content. Header shows 3,549 commits, 15.1k stars, 1.6k forks, no archive notice.

• https://raw.githubusercontent.com/modelscope/ms-swift/main/LICENSE recorded 2026-08-08

  11,357 bytes of unmodified Apache License 2.0 text: opens 'Apache License / Version 2.0, January 2004 / http://www.apache.org/licenses/' and closes with the stock appendix carrying the unfilled placeholder 'Copyright [yyyy] [name of copyright owner]'. No custom copyright line, no added use restriction, no field-of-use rider, so the OSI call is clean rather than resting on the GitHub classifier.

• https://swift.readthedocs.io/en/latest/ recorded 2026-08-08

  Full public documentation table of contents: Get Started (SWIFT Installation with 'Source Code Installation', Quick Start, Web-UI); Instruction (Command Line Parameters, Pre-training and Fine-tuning, GRPO, Knowledge Distillation, RLHF, Inference and Deployment, Sampling, Evaluation, Export and Push, Ray Support, Reinforced Fine-Tuning, Agent Support, Supported Models and Datasets, Using Tuners, FAQ); Megatron-SWIFT; Customization; Best Practices. A case-insensitive scan of the rendered page returns zero occurrences of 'enterprise', 'premium', 'pricing', 'commercial' or 'paid' - every documented capability is documented for the open package, with no tier withheld.

• https://pypi.org/project/ms-swift/ recorded 2026-08-08

  Project page for 'ms-swift 4.4.2', summary 'Swift: Scalable lightWeight Infrastructure for Fine-Tuning', author 'DAMO ModelScope teams'. Meta block gives License: 'Apache Software License (Apache License 2.0)' and the Classifiers block lists 'License :: OSI Approved :: Apache Software License'. The rendered long description is the repository README verbatim (same Table of Contents, same License section linking modelscope/ms-swift/blob/master/LICENSE), so the distributed package is the same framework the repo publishes.

122,381 PyPI downloads in the trailing 30 days for ms-swift, the project's only published package, which puts reach in the 100K-1M band.

• https://pypistats.org/api/packages/ms-swift/recent recorded 2026-08-13

  122,381 downloads in the trailing 30 days for ms-swift

Very broad coverage of both methods and models -- full fine-tuning, LoRA, QLoRA and DoRA, SFT, the GRPO family (GRPO, DAPO, GSPO, SAPO, CISPO, CHORD, RLOO, REINFORCE++), preference methods (DPO, KTO, reward modeling, CPO, SimPO, ORPO) and pretraining, over 600+ models -- with quantization, evaluation and deployment built into one pipeline. It sits one band below verl and Megatron-LM because it is a comprehensive practitioner toolkit rather than a demonstrated large-scale frontier trainer.

• https://github.com/modelscope/ms-swift recorded 2026-08-08

  README 'Supported Training Methods' matrix rows - Pre-training, Supervised Fine-Tuning, GRPO, GKD, PPO, DPO, KTO, Reward Model, CPO, SimPO, ORPO, Embedding, Reranker, Sequence Classification - across Full-Parameter / LoRA / QLoRA / Deepspeed / Multi-Machine / Multimodal columns. Separate Reinforcement Learning table lists GRPO, DAPO, GSPO, SAPO, CISPO, CHORD, RLOO, Reinforce++. Megatron-SWIFT table adds TP/PP/SP/CP/ETP/EP/VPP parallelism with FP8. Lightweight tuning list gives LoRA, QLoRA, DoRA, LoRA+, LLaMAPro, LongLoRA, LoRA-GA, ReFT, RS-LoRA, Adapter, LISA. Quantization export covers AWQ, GPTQ, FP8, BNB; evaluation runs through EvalScope over 100+ datasets; deployment through vLLM, SGLang and LmDeploy with an OpenAI-compatible interface.

• https://swift.readthedocs.io/en/latest/ recorded 2026-08-08

  Documentation sections corroborating the feature matrix with worked workflows rather than a claim: 'Pre-training and Fine-tuning' (Environment Preparation, Pre-training, Fine-tuning, Merge LoRA, Inference, Deployment), 'GRPO', 'Knowledge Distillation', 'RLHF' with sub-sections GRPO, DPO, RM, PPO, KTO, CPO, ORPO, SimPO, plus 'Sampling', 'Evaluation', 'Export and Push' (Merge LoRA, Quantization, Push Model), 'Ray Support' (Megatron Ray, Swift Ray), 'Reinforced Fine-Tuning', 'Agent Support', and a Megatron-SWIFT section covering LoRA Training, Multimodal Models, Mcore Bridge, Megatron GRPO, GKD and Ascend NPU.