AI Frameworks · modelscope

ms-swift

ms-swift is an open-source fine-tuning and deployment framework for 600+ large language models and 300+ multimodal models. It provides training (pre-training, SFT, DPO, GRPO), inference, quantization, and evaluation in a single toolkit with support for various hardware and lightweight training methods like LoRA.

Talk to DEV.co Discuss Implementation

Source: GitHub — github.com/modelscope/ms-swift

14.7k

GitHub stars

1.5k

Forks

Python

Primary language

Apache-2.0

License (OSI-approved)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

Field	Value
Repository	modelscope/ms-swift
Owner	modelscope
Primary language	Python
License	Apache-2.0 — OSI-approved
Stars	14.7k
Forks	1.5k
Open issues	696
Latest release	v4.4.0 (2026-07-06)
Last updated	2026-07-07
Source	https://github.com/modelscope/ms-swift

What ms-swift is

ms-swift integrates distributed training (DDP, DeepSpeed ZeRO, FSDP, Megatron TP/PP/CP/EP), memory optimization (GaLore, Flash-Attention, Ulysses/Ring-Attention), quantized training (BNB, AWQ, GPTQ), and reinforcement learning algorithms (GRPO family: GRPO, DAPO, GSPO, SAPO, CISPO, CHORD, RLOO). Supports multimodal packing, multiple inference engines (vLLM, SGLang, LMDeploy), and 150+ built-in datasets.

Quickstart

Get the ms-swift source

Clone the repository and explore it locally.

terminalbash

1git clone https://github.com/modelscope/ms-swift.git2cd ms-swift3# follow the project's README for install & configuration

Need it deployed, integrated, or customized instead? DEV.co ships production installs.

Best use cases

Cost-effective fine-tuning of LLMs and MLLMs at scale

Use LoRA, QLoRA, DoRA on quantized models (9GB for 7B models) with distributed training to fine-tune 600+ text and 300+ multimodal models with day-0 support for latest architectures (Qwen3, DeepSeek-V4, GLM-5.1, Llama4).

Reinforcement learning training with GRPO family algorithms

Build preference-optimized models using built-in GRPO, DAPO, GSPO, SAPO, CISPO, CHORD, RLOO with extensible reward functions, vLLM inference acceleration, and multi-turn scheduling for complex RL workflows.

Full-pipeline model deployment and evaluation

Train, quantize, benchmark, and serve models end-to-end using integrated evaluation (EvalScope, 100+ datasets), quantization export (AWQ, GPTQ, FP8), and inference acceleration (vLLM, SGLang, LMDeploy) with OpenAI API interface.

Implementation considerations

Requires Python 3.12, PyTorch ≥2.0, modelscope ≥1.23; verify dependency compatibility with your infrastructure before deployment.
Distributed training (Megatron, FSDP, DeepSpeed) requires careful cluster setup; single-machine LoRA training is simpler but may not achieve stated speedups.
Multi-modal packing and sequence parallelism (Ulysses/Ring-Attention) significantly boost throughput but add complexity; validate on representative datasets before production.
GRPO training requires reward model implementation or integration; reward function extensibility is noted but specifics on custom reward plugins require documentation review.
Integration with 150+ datasets implies schema variability; custom dataset formatting follows templates but expect pre-processing work for domain-specific data.

When to avoid it — and what to weigh

Proprietary/closed-source model training required — ms-swift is purpose-built for open-source and community models; not designed for training proprietary architectures outside the supported model lists.
Limited engineering resources or ML expertise — Framework requires substantial MLOps knowledge: distributed training setup, quantization tuning, GRPO hyperparameter optimization. Web-UI exists but core workflows demand deep configuration understanding.
Sub-9GB memory constraints on consumer hardware — Even with quantization and LoRA, minimal tested configurations assume ~9GB (7B models); training smaller models or very constrained setups not clearly documented.
Real-time or edge inference optimization as primary goal — Framework emphasizes training and cloud/server deployment; edge optimization (TensorRT, CoreML, ONNX export) not prominently featured.

License & commercial use

Apache License 2.0 (Apache-2.0). Permissive OSI-approved license allowing commercial use, modification, and distribution with attribution and liability disclaimer.

Apache-2.0 is a permissive license that explicitly permits commercial use. You may use ms-swift to build and deploy commercial services, fine-tune proprietary data, and distribute derivatives provided you retain license notices and include a copy of the Apache-2.0 license. No commercial license restrictions apply; no royalties required. However, any derivative code modifications remain subject to Apache-2.0 terms.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

Signal	Assessment
Maintenance	Active
Documentation	Adequate
License clarity	Clear
Deployment complexity	High
DEV.co fit	Good
Assessment confidence	High

Security considerations

No security vulnerabilities or posture claims can be validated from this data. Standard considerations: (1) Model artifacts and training data security—ensure data privacy for fine-tuning (credentials, inference logs); (2) Supply chain—vet community datasets and model checkpoints for provenance; (3) Quantization/inference—ensure deployed models undergo threat modeling if used in sensitive contexts; (4) Dependency management—keep PyTorch, transformers, and auxiliary packages patched. Security audit status unknown; use in high-security environments warrants threat modeling review.

Alternatives to consider

Hugging Face Transformers + TRL/SFTTrainer

Lighter-weight, more modular stack; narrower model support (fewer LLMs/MLLMs Day-0); strong community ecosystem but requires more manual integration for distributed training and quantization pipelines.

LitGPT (Lightning AI)

Simpler, single-machine focused; supports fewer models; easier for small teams but lacks Megatron parallelism, GRPO algorithms, and multimodal training at ms-swift's scale.

Ollama + local quantization

Extremely lightweight and beginner-friendly for inference/deployment; not designed for training or fine-tuning; no distributed/RL capabilities; suitable for inference-only use cases.

Software development agency

Build on ms-swift with DEV.co software developers

ms-swift provides day-0 support for latest models (Qwen3, DeepSeek-V4, Llama4) with lightweight LoRA or full-parameter training, distributed parallelism, and production deployment. Evaluate your use case against complexity and integration needs, then explore examples on GitHub or join the Discord community.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

ml-engineering ludwig axolotl

Related on DEV.co

Explore the category and the services that help you build with it.

AI Frameworks AI development API development Cloud deployment Talk to DEV.co

ms-swift FAQ

What is the minimum hardware to run ms-swift fine-tuning?

~9GB VRAM for 7B models using QLoRA on quantized models. Supports A10/A100/H100, RTX series, T4/V100, AMD MI300 series, Ascend NPU, and CPU training (slower). GPU strongly recommended for practical throughput.

Does ms-swift support custom models outside the 600+ LLM/300+ MLLM list?

Not clearly stated in provided data. Framework targets community/open-source models with Day-0 support for popular architectures. Custom model integration likely requires code extension; specifics require documentation review.

What is the time-to-training-ready for a new project?

Minimal: CLI commands run fine-tuning on pre-built datasets in hours on moderate hardware. Significant: custom datasets, reward models (GRPO), or distributed clusters add days of setup, tuning, and validation.

Can I use ms-swift models in production at scale?

Yes, with proper infrastructure. Inference acceleration via vLLM/SGLang/LMDeploy enables production serving. Deployment orchestration (Kubernetes, monitoring, auth) is user's responsibility; not provided by framework.

Software development & web development with DEV.co

From first prototype to production, DEV.co delivers software development services around tools like ms-swift. Our software development agency staffs experienced software developers and web developers for custom software development, web development, integrations, and ongoing support across ai frameworks and beyond.

Ready to Fine-Tune Models at Scale?

Talk to DEV.co Schedule a Consultation