LLM/MLLM fine-tuning and deployment
ms-swift is a Python framework from ModelScope for fine-tuning, alignment, inference, evaluation, quantization and deployment of hundreds of LLMs and MLLMs.
Same problem, same approach. Swapping one for another is a config change, not a rewrite.
Choose when you prefer an NVIDIA-maintained PyTorch-native distributed training library with direct HuggingFace model support.
🚀 Pytorch Distributed native training library for LLMs/VLMs with OOTB Hugging Face support
Choose when you want a HuggingFace-centric unified SFT/DPO/GRPO harness with a broader plug-in ecosystem and a different UI workflow.
Unified Efficient Fine-Tuning of 100+ LLMs & VLMs (ACL 2024)
Choose when you want a reproducible, YAML-config-driven LLM fine-tune/evaluate/deploy framework independent of the ModelScope ecosystem.
Easily fine-tune, evaluate and deploy Qwen, Gemma, or any open weight LLM!
Solves the same problem with a different architecture or at a different layer. Expect to rewrite the integration.
Choose when you want a lightweight Lightning/Fabric-based pretrain-finetune-serve pipeline for a curated set of LLMs rather than the huge ms-swift model zoo.
20+ high-performance LLMs with recipes to pretrain, finetune and deploy at scale.
Choose when your main goal is maximum-speed, low-VRAM LoRA fine-tuning on local/consumer GPUs and you do not need the full fine-tune/deploy lifecycle.
Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4, MiniMax-H3, Gemma 4, FLUX and more.
Choose when you are working in JAX/TPU or large-scale GPU clusters and want scalable training, accepting a wholly different stack from ms-swift.
A simple, performant, and scalable Jax LLM!
Choose when you want a training engine specialized for ultra-large MoE/InternLM-scale models and plan to deploy through the LMDeploy ecosystem.
A Next-Generation Training Engine Built for Ultra-Large MoE Models
An earlier or more limited way to do the job. Still the right call when you need something small, proven, or CPU-only.
Choose when you must fine-tune on a very low-VRAM laptop GPU using layer streaming, accepting much narrower model/algorithm coverage.
Fine-tune LLMs from one YAML. Layer streaming trains an 8B model on a 4 GB laptop GPU.
Choose only for a simpler, older-style LLM fine-tuning library with fewer modern model families and no comparable MLLM/deployment support.
Build, personalize and control your own LLMs. From data pre-processing to fine-tuning, xTuring provides an easy way to personalize open-source LLMs. Join our discord community: https://discord.gg/TgHXuSJEk6
Choose when you specifically need light omni-modal reinforcement fine-tuning with reward models; it is a narrow subset of ms-swift's alignment options.
LightRFT: Light, Efficient, Omni-modal & Reward-model Driven Reinforcement Fine-Tuning Framework
These projects were analysed and named ms-swift among their alternatives. The relationship is not symmetric — how ms-swift rates them is a separate judgement, made when ms-swift is analysed in its own right.
calls ms-swift “choose when you need even broader model coverage (600+ LLMs, 300+ MLLMs) with tight ModelScope ecosystem integration.”
Unified Efficient Fine-Tuning of 100+ LLMs & VLMs (ACL 2024)
calls ms-swift “Choose it when out-of-the-box coverage of 600+ LLMs and 300+ MLLMs matters more than Oumi's integrated evaluation and deployment tools.”
Easily fine-tune, evaluate and deploy Qwen, Gemma, or any open weight LLM!
calls ms-swift “Choose for extremely broad LLM/MLLM support (600+ models) with a Python/CLI training workflow.”
Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4, MiniMax-H3, Gemma 4, FLUX and more.
calls ms-swift “PyTorch/ModelScope suite for SFT/DPO/GRPO across many LLMs; choose for non-Paddle training workflows.”
PaddleFormers is an easy-to-use library of pre-trained large language model zoo based on PaddlePaddle.