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

Oumi is a fully open-source platform that simplifies the entire lifecycle of foundation models: data preparation, training, evaluation, and deployment. Whether you're experimenting on a laptop or running large-scale training on a cluster, Oumi provides the tools and workflows you need.

✅ Train & Fine-Tune models from 10M to 405B parameters (LoRA, QLoRA, DPO, SFT)
🤖 Work with Multimodal Models (Llama, DeepSeek, Qwen, Phi, and more)
📊 Evaluate Models using comprehensive benchmarks
🔄 Synthesize & Curate Data with LLM-powered judges
⚡ Deploy Models Efficiently with optimized inference engines (vLLM, SGLang)
🌎 Run Anywhere: Laptops, Clusters, or Cloud (AWS, Azure, GCP, Lambda)
🔌 Integrate Seamlessly with OpenAI, Anthropic, Vertex AI, Together, Parasail, and more

All with one consistent API, production-grade reliability, and research flexibility.

# Install Oumi (CPU & NPU only)
pip install oumi  

# OR, with GPU support (Requires Nvidia or AMD GPU)
pip install oumi[gpu]  

# Install the latest version from source
git clone https://github.com/oumi-ai/oumi.git
cd oumi
pip install .

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

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

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

For more details, check the installation guide.

Easily run your jobs on major cloud platforms:

# 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_aws_job.yaml

# Azure
oumi launch up -c configs/recipes/smollm/sft/135m/quickstart_azure_job.yaml

# Lambda
oumi launch up -c configs/recipes/smollm/sft/135m/quickstart_lambda_job.yaml

Note: Oumi is in beta, and some features are still evolving.

• Ready-to-use recipes for popular models & workflows
• No need to write training loops or data pipelines

• Built for scalability and reliability
• Proven by teams training models at scale

• Easily reproducible experiments
• Flexible interfaces for customizing every component

• Supports tiny to large-scale models, including multimodal ones
• Compatible with Hugging Face Transformers

• Built-in support for distributed training (FSDP, DDP)
• Optimized inference engines (vLLM, SGLang)

• 100% open-source with no vendor lock-in
• Active developer and research community

Oumi supports a variety of foundation models with ready-to-use configurations:

For a full list of supported models, check the Oumi documentation.

We welcome contributions! To get started:

1. Read the CONTRIBUTING.md
2. Join our Discord community to discuss, ask questions, and share insights.
3. Explore open issues and start contributing 🚀

Oumi is licensed under the Apache License 2.0. See the LICENSE file for more details.

If you use Oumi in your research, please cite it:

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

Oumi is built on the shoulders of giants! A huge thanks to the open-source community and contributors who make this possible. 🚀