oumi-ai/oumi：轻松微调（SFT/RL）、评估并部署 Qwen、Gemma

技术解读

以下内容由 GitRead 中文社区编辑团队基于项目 README 与官方资料整理，结合中文开发者的常见使用场景补充技术选型建议。具体以项目官方文档为准。

项目背景与痛点

在当前的 AI 生态系统中，大语言模型（LLM）已经变得越来越普遍。无论是 Qwen 还是 Gemma，这些模型都展现了其强大的文本生成和处理能力。然而，要想将这些预先训练好的大语言模型真正应用于实际生产环境中，我们需要一个易于使用的框架来简化微调、评估和部署的过程。oumi 正是为了解决这一需求而诞生的工具。它面向广大开发者和企业用户，致力于提供一个直观且高效的环境，让使用和优化大规模语言模型变得更加简单。

核心技术揭秘

oumi 的核心技术在于其精简但功能强大的架构设计。通过搭配包括Python在内的多种语言工具链，以及一套清晰的 API 接口，oumi 能够无缝对接各种语言模型，进一步帮助企业开发者在实际生产环境中快速部署和优化语言模型。

graph TD A[用户请求] --> B{"参数校验"} B -->|"通过"| C[执行业务] C --> D[调用语言模型微调] D --> E{"评估结果"} E -->|"达标"| F[部署生产环境] E -->|"未达标"| G[调整模型参数] G --> H[模型调整] H --> I[重新评估] I -->|"仍不达标"| G I -->|"达标"| F

透过上述架构可以清晰地看到用户请求如何经过参数校验，随后进入具体的业务执行阶段。调用语言模型微调环节是关键部分，利用 Python 和其他工具链进行模型优化后，评估结果是否达标。若不达标，将调整模型参数并再次进行评估；只有达标后，模型才会被部署到生产环境中。

功能亮点与差异

oumi 的功能亮点在于其面向领域的优化和高度灵活的配置能力。与竞品相比，oumi 提供更多的微调选项以及更完善的评估系统，能够通过自定义规则来优化模型性能。灵活的配置能力加上强大的社区支持使得 oumi 相比竞品更具有吸引力。

应用场景与落地建议

在具体使用场景中，开发者可将 oumi 集成到文档写作、自动问答系统、文本生成等业务场景中。在生产环境中使用 oumi 时需要注意以下几点：

参数校验至关重要，务必做好输入数据的预处理；
模型微调过程耗时，尤其是在大规模数据集上；
定期评估模型性能并及时调整，以持续优化模型效果。

综合评价

oumi 作为一款开源工具，能够在很大程度上简化使用和部署大规模语言模型的过程。其胜在灵活配置和开源的社区支持，但也需要注意参数校验的重要性以及模型微调和评估的效率问题。总的来说，对于希望快速部署和优化语言模型的开发者和企业用户来说，oumi 是一种值得考虑的高效工具。

关于

GitHub 原始描述（英文）

Easily fine-tune (SFT/RL), evaluate, and deploy Qwen, Gemma, or any open agentic LLM/VLM!

agentic-ai dpo fine-tuning gemma gpt-oss grpo inference llama llm llm-as-a-judge llm-evaluation llms lora pytorch qwen reinforcement-learning sft synthetic-data tool-calling vlm

Apache License 2.0
9.4k Stars
798 Forks
61 Watching

描述取自 GitHub 官方 API，保持仓库作者书写的原文。

简要分析

热度分

11088

价值分

5902

活跃状态

活跃

GitHub

语言占比

Dockerfile

Jinja

Jupyter Notebook

Makefile

Python

Shell

Release

v0.99月17日
v0.85月8日
v0.71月30日
v0.6.02025年12月18日
v0.5.02025年11月19日

贡献者

Everything you need to build state-of-the-art foundation models, end-to-end

🔥 News

[2026/09] Oumi v0.9 released: an end-to-end stack for agentic, tool-using models, with tool-calling SFT/DPO, executable tool environments, GRPO over any Oumi environment with verl, and a multi-criteria RubricJudge.
[2026/08] Extended GRPO reinforcement learning to support tool use
[2026/07] Added support for tools, environments (simulated, lookup, database), and agentic data synthesis
[2026/06] Added support for the Gemma 4 model family
[2026/06] Added partial-failure support across inference, judging, and data synthesis
[2026/05] Oumi v0.8 released with oumi deploy CLI for dedicated inference endpoints, an oumi-mcp MCP server for Claude/Cursor integration, batch API support across Anthropic/Fireworks/Together, and Transformers v5 / TRL / vLLM dependency upgrades
[2026/03] Upgraded to Transformers v5, TRL v0.30, vLLM v0.19, and veRL v0.7 compatibility
[2026/03] New oumi deploy command for deploying Oumi models to dedicated inference endpoints on Fireworks and Parasail
[2026/03] Added support for the Qwen3.5 model family
[2026/02] Preview of using the Oumi Platform and Lambda to fine-tune and deploy a 4B model for user intent classification
[2026/02] Lambda and Oumi partner for end-to-end custom model development
[2026/01] Oumi v0.7 released with Python 3.14 and PyTorch 2.9 support, Fireworks and OpenRouter inference engines, and rule-based evaluation judges
[2025/12] Oumi v0.6.0 released with Python 3.13 support, oumi analyze CLI command, TRL 0.26+ support, and more
[2025/12] WeMakeDevs AI Agents Assemble Hackathon: Oumi webinar on Finetuning for Text-to-SQL
[2025/12] Oumi co-sponsors WeMakeDevs AI Agents Assemble Hackathon with over 2000 project submissions
[2025/11] Oumi v0.5.0 released with advanced data synthesis, hyperparameter tuning automation, support for OpenEnv, and more

Older updates

[2025/11] Example notebook to perform RLVF fine-tuning with OpenEnv, an open source library from the Meta PyTorch team for creating, deploying, and distributing agentic RL environments
[2025/10] Oumi v0.4.1 and v0.4.2 released] with support for Qwen3-VL and Transformers v4.56, data synthesis documentation and examples, and many bug fixes
[2025/09] Oumi v0.4.0 released with DeepSpeed support, a Hugging Face Hub cache management tool, KTO/Vision DPO trainer support
[2025/08] Training and inference support for OpenAI’s gpt-oss-20b and gpt-oss-120b: recipes here
[2025/08] Aug 14 Webinar - OpenAI’s gpt-oss: Separating the Substance from the Hype.
[2025/08] Oumi v0.3.0 released with model quantization (AWQ), an improved LLM-as-a-Judge API, and Adaptive Inference
[2025/07] Recipe for Qwen3 235B
[2025/07] July 24 webinar: “Training a State-of-the-art Agent LLM with Oumi + Lambda”
[2025/06] Oumi v0.2.0 released with support for GRPO fine-tuning, a plethora of new model support, and much more
[2025/06] Announcement of Data Curation for Vision Language Models (DCVLR) competition at NeurIPS2025
[2025/06] Recipes for training, inference, and eval with the newly released Falcon-H1 and Falcon-E models
[2025/05] Support and recipes for InternVL3 1B
[2025/04] Added support for training and inference with Llama 4 models: Scout (17B activated, 109B total) and Maverick (17B activated, 400B total) variants, including full fine-tuning, LoRA, and QLoRA configurations
[2025/04] Recipes for Qwen3 model family
[2025/04] Introducing HallOumi: a State-of-the-Art Claim-Verification Model (technical overview)
[2025/04] Oumi now supports two new Vision-Language models: Phi4 and Qwen 2.5

🔎 About

Oumi is a fully open-source platform that streamlines the entire lifecycle of foundation models - from data preparation and training to evaluation and deployment. Whether you’re developing on a laptop, launching large scale experiments on a cluster, or deploying models in production, Oumi provides the tools and workflows you need.

With Oumi, you can:

🚀 Train and fine-tune models from 10M to 400B+ parameters using state-of-the-art techniques (SFT, LoRA, QLoRA, DPO, GRPO, and more)
🛠️ Train agentic, tool-using models: tool-calling SFT, executable tool environments (database, HTTP endpoint, lookup, simulated), and reinforcement learning over those environments with verl GRPO
🤖 Work with both text and multimodal models (Llama, Gemma, Qwen, DeepSeek, GLM, Phi, and others)
🔄 Synthesize and curate training data, including multi-turn tool-use conversations, with LLM judges
⚡️ Serve models efficiently with popular inference engines (vLLM, SGLang), or deploy to dedicated endpoints with oumi deploy
📊 Evaluate models across standard benchmarks and with multi-criteria rubric judges
🌎 Run anywhere - from laptops to clusters to clouds (AWS, Azure, GCP, Lambda, Modal, Slurm, and more)
🔌 Integrate with both open models and commercial APIs (OpenAI, Anthropic, Vertex AI, Together, Fireworks, Parasail, …), and drive Oumi from Claude or Cursor with the oumi-mcp MCP server

All with one consistent API, production-grade reliability, and all the flexibility you need for research.

Learn more at oumi.ai, or jump right in with the quickstart guide.

🚀 Getting Started

Notebook	Try in Colab	Goal
🎯 Getting Started: A Tour		Quick tour of core features: training, evaluation, inference, and job management
🔧 Model Finetuning Guide		End-to-end guide to LoRA tuning with data prep, training, and evaluation
📚 Model Distillation		Guide to distilling large models into smaller, efficient ones
📋 Model Evaluation		Comprehensive model evaluation using Oumi’s evaluation framework
☁️ Remote Training		Launch and monitor training jobs on cloud (AWS, Azure, GCP, Lambda, etc.) platforms
📈 LLM-as-a-Judge		Filter and curate training data with built-in judges
🎯 GRPO Reinforcement Learning		Train a model with GRPO and a custom reward function
🖼️ Vision Language Models		Fine-tune and run inference with vision-language models
⚡️ vLLM Inference		Run fast local inference with the vLLM engine

🔧 Usage

Installation

Choose the installation method that works best for you:

Using pip (Recommended)

# Basic installation
uv pip install oumi

# With GPU support
uv pip install 'oumi[gpu]'

# Latest development version
uv pip install git+https://github.com/oumi-ai/oumi.git

Don’t have uv? Install it or use pip instead.

Using Docker

# Pull the latest image
docker pull ghcr.io/oumi-ai/oumi:latest

# Run oumi commands
docker run --gpus all -it ghcr.io/oumi-ai/oumi:latest oumi --help

# Train with a mounted config
docker run --gpus all -v $(pwd):/workspace -it ghcr.io/oumi-ai/oumi:latest \
    oumi train --config /workspace/my_config.yaml

Quick Install Script (Experimental)

Try Oumi without setting up a Python environment. This installs Oumi in an isolated environment:

curl -LsSf https://oumi.ai/install.sh | bash

For more advanced installation options, see the installation guide.

Oumi CLI

You can quickly use the oumi command to train, evaluate, and infer models using one of the existing recipes:

# Training
oumi train -c configs/recipes/smollm/sft/135m/quickstart_train.yaml

# Evaluation
oumi evaluate -c configs/recipes/smollm/evaluation/135m/quickstart_eval.yaml

# Inference
oumi infer -c configs/recipes/smollm/inference/135m_infer.yaml --interactive

For more advanced options, see the training, evaluation, inference, and llm-as-a-judge guides.

Running Jobs Remotely

You can run jobs remotely on cloud platforms (AWS, Azure, GCP, Lambda, etc.) using the oumi launch command:

# GCP
oumi launch up -c configs/recipes/smollm/sft/135m/quickstart_gcp_job.yaml

# AWS
oumi launch up -c configs/recipes/smollm/sft/135m/quickstart_gcp_job.yaml --resources.cloud aws

# Azure
oumi launch up -c configs/recipes/smollm/sft/135m/quickstart_gcp_job.yaml --resources.cloud azure

# Lambda
oumi launch up -c configs/recipes/smollm/sft/135m/quickstart_gcp_job.yaml --resources.cloud lambda

Note: Oumi is in beta and under active development. The core features are stable, but some advanced features might change as the platform improves.

💻 Why use Oumi?

If you need a comprehensive platform for training, evaluating, or deploying models, Oumi is a great choice.

Here are some of the key features that make Oumi stand out:

🔧 Zero Boilerplate: Get started in minutes with ready-to-use recipes for popular models and workflows. No need to write training loops or data pipelines.
🏢 Enterprise-Grade: Built and validated by teams training models at scale
🎯 Research Ready: Perfect for ML research with easily reproducible experiments, and flexible interfaces for customizing each component.
🌐 Broad Model Support: Works with most popular model architectures - from tiny models to the largest ones, text-only to multimodal.
🚀 SOTA Performance: Native support for distributed training techniques (FSDP, DeepSpeed, DDP) and optimized inference engines (vLLM, SGLang).
🤝 Community First: 100% open source with an active community. No vendor lock-in, no strings attached.

📚 Examples & Recipes

Explore the growing collection of ready-to-use configurations for state-of-the-art models and training workflows:

Note: These configurations are not an exhaustive list of what’s supported, simply examples to get you started. You can find a more exhaustive list of supported models, and datasets (supervised fine-tuning, pre-training, preference tuning, and vision-language finetuning) in the oumi documentation.

Qwen Family

Model	Example Configurations
Qwen3-Next 80B A3B	LoRA • Inference • Inference (Instruct) • Evaluation
Qwen3.5 0.8B	FFT • LoRA • Inference (vLLM) • Inference
Qwen3 30B A3B	LoRA • Inference • Evaluation
Qwen3 32B	LoRA • Inference • Evaluation
Qwen3 14B	LoRA • Inference • Evaluation
Qwen3 8B	FFT • Inference • Evaluation
Qwen3 4B	FFT • Inference • Evaluation
Qwen3 1.7B	FFT • Inference • Evaluation
Qwen3 0.6B	FFT • Inference • Evaluation
Qwen3-Coder 30B A3B Instruct	Inference (vLLM) • Inference (vLLM, tool calling) • Inference (GGUF) • Inference (GGUF, macOS)
QwQ 32B	FFT • LoRA • QLoRA • Inference • Evaluation
Qwen2.5-VL 3B	SFT • LoRA • Inference (vLLM) • Inference
Qwen2-VL 2B	SFT • LoRA • Inference (vLLM) • Inference (SGLang) • Inference • Evaluation

💎 Gemma Family

Model	Example Configurations
Gemma 4 E2B Instruct	FFT • LoRA
Gemma 4 E4B Instruct	FFT • LoRA
Gemma 4 12B Instruct	FFT • LoRA
Gemma 4 26B A4B Instruct	LoRA
Gemma 4 31B Instruct	LoRA
Gemma 3 4B Instruct	FFT • Inference • Evaluation
Gemma 3 12B Instruct	LoRA • Inference • Evaluation
Gemma 3 27B Instruct	LoRA • Inference • Evaluation

🐋 DeepSeek R1 Family

Model	Example Configurations
DeepSeek R1 671B	Inference (Together AI)
Distilled Llama 8B	FFT • LoRA • QLoRA • Inference • Evaluation
Distilled Llama 70B	FFT • LoRA • QLoRA • Inference • Evaluation
Distilled Qwen 1.5B	FFT • LoRA • Inference • Evaluation
Distilled Qwen 32B	LoRA • Inference • Evaluation

🦙 Llama Family

Model	Example Configurations
Llama 4 Scout Instruct 17B	FFT • LoRA • QLoRA • Inference (vLLM) • Inference • Inference (Together.ai)
Llama 4 Scout 17B	FFT
Llama 3.1 8B	FFT • LoRA • QLoRA • Pre-training • Inference (vLLM) • Inference • Evaluation
Llama 3.1 70B	FFT • LoRA • QLoRA • Inference • Evaluation
Llama 3.1 405B	FFT • LoRA • QLoRA
Llama 3.2 1B	FFT • Inference (vLLM) • Inference (SGLang) • Inference • Evaluation
Llama 3.2 3B	FFT • LoRA • QLoRA • Inference (vLLM) • Inference (SGLang) • Inference • Evaluation
Llama 3.3 70B	FFT • LoRA • QLoRA • Inference (vLLM) • Inference • Evaluation
Llama 3.2 Vision 11B	SFT • Inference (vLLM) • Inference (SGLang) • Evaluation

🦅 Falcon family

Model	Example Configurations
Falcon-H1	FFT • Inference • Evaluation
Falcon-E (BitNet)	FFT • DPO • Evaluation

🧠 gpt-oss Family

Model	Example Configurations
gpt-oss 20B	LoRA (single GPU) • LoRA (multi-GPU) • Inference (vLLM) • Inference
gpt-oss 120B	LoRA (multi-GPU) • Inference (vLLM) • Inference (Together AI) • Inference

🧩 GLM Family

Model	Example Configurations
GLM-4.7	Inference (Fireworks)
GLM-4.5-Air	Inference (GGUF) • Inference (GGUF, macOS)
GLM-4 9B Chat	Inference (vLLM)

🦉 OLMo 3 Family

Model	Example Configurations
OLMo 3 7B Instruct	FFT • Inference • Evaluation
OLMo 3 32B Instruct	LoRA • Inference • Evaluation

🎨 Vision Models

Model	Example Configurations
Llama 3.2 Vision 11B	SFT • LoRA • Inference (vLLM) • Inference (SGLang) • Evaluation
LLaVA 7B	SFT • Inference (vLLM) • Inference
Phi3 Vision 4.2B	SFT • LoRA • Inference (vLLM)
Phi4 Vision 5.6B	SFT • LoRA • Inference (vLLM) • Inference
Qwen2-VL 2B	SFT • LoRA • Inference (vLLM) • Inference (SGLang) • Inference • Evaluation
InternVL3 1B	SFT
Molmo 7B-D	SFT
Molmo 7B-O	SFT • GRPO
Qwen3-VL 2B	Inference
Qwen3-VL 4B	Inference
Qwen3-VL 8B	Inference
Qwen3-VL 30B A3B Instruct	FFT • LoRA
Qwen2.5-VL 3B	SFT • LoRA • Inference (vLLM) • Inference
SmolVLM-Instruct 2B	SFT • LoRA

🔍 Even more options

This section lists all the language models that can be used with Oumi. Thanks to the integration with the 🤗 Transformers library, you can easily use any of these models for training, evaluation, or inference.

Models prefixed with a checkmark (✅) have been thoroughly tested and validated by the Oumi community, with ready-to-use recipes available in the configs/recipes directory.

📋 Click to see more supported models

Instruct Models

Model	Size	Paper	HF Hub	License	Open [^1]
✅ SmolLM-Instruct	135M/360M/1.7B	Blog	Hub	Apache 2.0	✅
✅ DeepSeek R1 Family	1.5B/8B/32B/70B/671B	Blog	Hub	MIT	❌
✅ Llama 3.1 Instruct	8B/70B/405B	Paper	Hub	License	❌
✅ Llama 3.2 Instruct	1B/3B	Paper	Hub	License	❌
✅ Llama 3.3 Instruct	70B	Paper	Hub	License	❌
✅ Phi-3.5-Instruct	4B/14B	Paper	Hub	License	❌
✅ Qwen3	0.6B-32B	Paper	Hub	License	❌
Qwen2.5-Instruct	0.5B-70B	Paper	Hub	License	❌
OLMo 2 Instruct	7B	Paper	Hub	Apache 2.0	✅
✅ OLMo 3 Instruct	7B/32B	Paper	Hub	Apache 2.0	✅
MPT-Instruct	7B	Blog	Hub	Apache 2.0	✅
Command R	35B/104B	Blog	Hub	License	❌
Granite-3.1-Instruct	2B/8B	Paper	Hub	Apache 2.0	❌
Gemma 2 Instruct	2B/9B	Blog	Hub	License	❌
✅ Gemma 4 Instruct	E2B/E4B/12B/26B A4B/31B	Blog	Hub	License	❌
✅ Qwen3.5	0.8B	Blog	Hub	License	❌
GLM-4.5-Air	106B (Total)	Blog	Hub	MIT	❌
✅ Gemma 3 Instruct	4B/12B/27B	Blog	Hub	License	❌
DBRX-Instruct	130B MoE	Blog	Hub	Apache 2.0	❌
Falcon-Instruct	7B/40B	Paper	Hub	Apache 2.0	❌
✅ Llama 4 Scout Instruct	17B (Activated) 109B (Total)	Paper	Hub	License	❌
✅ Llama 4 Maverick Instruct	17B (Activated) 400B (Total)	Paper	Hub	License	❌

Vision-Language Models

Model	Size	Paper	HF Hub	License	Open
✅ Llama 3.2 Vision	11B	Paper	Hub	License	❌
✅ LLaVA-1.5	7B	Paper	Hub	License	❌
✅ Phi-3 Vision	4.2B	Paper	Hub	License	❌
✅ BLIP-2	3.6B	Paper	Hub	MIT	❌
✅ Qwen2-VL	2B	Blog	Hub	License	❌
✅ Qwen3-VL	2B/4B/8B	Blog	Hub	License	❌
✅ SmolVLM-Instruct	2B	Blog	Hub	Apache 2.0	✅

Base Models

Model	Size	Paper	HF Hub	License	Open
✅ SmolLM2	135M/360M/1.7B	Blog	Hub	Apache 2.0	✅
✅ Llama 3.2	1B/3B	Paper	Hub	License	❌
✅ Llama 3.1	8B/70B/405B	Paper	Hub	License	❌
✅ GPT-2	124M-1.5B	Paper	Hub	MIT	✅
DeepSeek V2	7B/13B	Paper	Hub	License	❌
Gemma2	2B/9B	Blog	Hub	License	❌
GPT-J	6B	Blog	Hub	Apache 2.0	✅
GPT-NeoX	20B	Paper	Hub	Apache 2.0	✅
Mistral	7B	Paper	Hub	Apache 2.0	❌
Mixtral	8x7B/8x22B	Blog	Hub	Apache 2.0	❌
MPT	7B	Blog	Hub	Apache 2.0	✅
OLMo	1B/7B	Paper	Hub	Apache 2.0	✅
✅ Llama 4 Scout	17B (Activated) 109B (Total)	Paper	Hub	License	❌

Reasoning Models

Model	Size	Paper	HF Hub	License	Open
✅ gpt-oss	20B/120B	Paper	Hub	Apache 2.0	❌
✅ Qwen3	0.6B-32B	Paper	Hub	License	❌
✅ Qwen3-Next	80B-A3B	Blog	Hub	License	❌
Qwen QwQ	32B	Blog	Hub	License	❌

Code Models

Model	Size	Paper	HF Hub	License	Open
✅ Qwen2.5 Coder	0.5B-32B	Blog	Hub	License	❌
DeepSeek Coder	1.3B-33B	Paper	Hub	License	❌
StarCoder 2	3B/7B/15B	Paper	Hub	License	✅

Math Models

Model	Size	Paper	HF Hub	License	Open
DeepSeek Math	7B	Paper	Hub	License	❌

📖 Documentation

To learn more about all the platform’s capabilities, see the Oumi documentation.

🤝 Join the Community

Oumi is a community-first effort. Whether you are a developer, a researcher, or a non-technical user, all contributions are very welcome!

To contribute to the oumi repository, please check the CONTRIBUTING.md for guidance on how to contribute to send your first Pull Request.
Make sure to join our Discord community to get help, share your experiences, and contribute to the project!

🙏 Acknowledgements

Oumi makes use of several libraries and tools from the open-source community. We would like to acknowledge and deeply thank the contributors of these projects! ✨ 🌟 💫

📝 Citation

If you find Oumi useful in your research, please consider citing it:

@software{oumi2025,
  author = {Oumi Community},
  title = {Oumi: an Open, End-to-end Platform for Building Large Foundation Models},
  month = {January},
  year = {2025},
  url = {https://github.com/oumi-ai/oumi}
}

📜 License

This project is licensed under the Apache License 2.0. See the LICENSE file for details.

[^1]: Open models are defined as models with fully open weights, training code, and data, and a permissive license. See Open Source Definitions for more information.

暂无评论

oumi

技术解读

项目背景与痛点

核心技术揭秘

功能亮点与差异

应用场景与落地建议

综合评价

相关仓库

关于

简要分析

语言占比

Release

贡献者

Everything you need to build state-of-the-art foundation models, end-to-end

🔥 News

🔎 About

🚀 Getting Started

🔧 Usage

Installation

Oumi CLI

Running Jobs Remotely

💻 Why use Oumi?

📚 Examples & Recipes

Qwen Family

💎 Gemma Family

🐋 DeepSeek R1 Family

🦙 Llama Family

🦅 Falcon family

🧠 gpt-oss Family

🧩 GLM Family

🦉 OLMo 3 Family

🎨 Vision Models

🔍 Even more options

Instruct Models

Vision-Language Models

Base Models

Reasoning Models

Code Models

Math Models

📖 Documentation

🤝 Join the Community

🙏 Acknowledgements

📝 Citation

📜 License

评论