Open Universal Machine Intelligence
Everything you need to build state-of-the-art foundation models, end-to-end.
Get Started →What is Oumi?#
Oumi is an open-source platform designed for ML engineers and researchers who want to train, fine-tune, evaluate, and deploy foundation models. Whether you’re fine-tuning a small language model on a single GPU or training a 405B parameter model across a cluster, Oumi provides a unified interface that scales with your needs.
Who is Oumi for?
ML Engineers building production AI systems who need reliable training pipelines and deployment options
Researchers experimenting with new training methods, architectures, or datasets
Teams who want a consistent workflow from local development to cloud-scale training
What problems does Oumi solve?
Fragmented tooling: Instead of stitching together different libraries for training, evaluation, and deployment, Oumi provides one cohesive platform
Scaling complexity: The same configuration works locally and on cloud infrastructure (AWS, GCP, Azure, Lambda Labs)
Reproducibility: YAML-based configs make experiments easy to track, share, and reproduce
Quick Start#
Prerequisites: Python 3.10+, pip. GPU recommended for larger models (CPU works for small models like SmolLM-135M).
Install Oumi and start training in minutes:
# Install with GPU support (or use `pip install oumi` for CPU-only)
pip install oumi[gpu]
# Train a model
oumi train -c configs/recipes/smollm/sft/135m/quickstart_train.yaml
# Run inference
oumi infer -c configs/recipes/smollm/inference/135m_infer.yaml --interactive
For detailed setup instructions including virtual environments and cloud setup, see the installation guide.
What will you build?#
Oumi provides a unified interface across the entire model development lifecycle. The workflows below cover training, evaluation, inference, data synthesis, hyperparameter tuning, and cloud deployment—all driven by YAML configs that work identically on your laptop or a multi-node cluster.
Start with a pre-trained model and customize it for your task using SFT, LoRA, DPO, GRPO, and more.
Run benchmarks and compare against baselines using standard evaluation suites and LLM judges.
Run inference anywhere—vLLM and llama.cpp locally, or OpenAI and Anthropic remotely—with a unified interface.
Create high-quality training data with LLM-powered synthesis pipelines.
Find the best learning rate, batch size, and other settings automatically using bayesian optimization.
Launch jobs on AWS, GCP, Azure, or Lambda Labs with a single command.
Hands-on Notebooks#
Explore the most common Oumi workflows hands-on. These notebooks run in Google Colab with pre-configured environments—just click and start experimenting. Try “A Tour” for a high-level overview, or dive straight into a specific topic.
Quick tour of core features: training, evaluation, inference, and job management
End-to-end guide to LoRA tuning with data prep, training, and evaluation
Guide to distilling large models into smaller, efficient ones
Comprehensive model evaluation using Oumi’s evaluation framework
Launch and monitor training jobs on cloud platforms (AWS, Azure, GCP, Lambda)
Filter and curate training data with built-in judges
Community & Support#
Oumi is a community-first effort. Whether you are a developer, a researcher, or a non-technical user, all contributions are very welcome!
Join our Discord community to get help, share your experiences, and chat with the team
Check the FAQ for common questions and troubleshooting
Open an issue on GitHub for bug reports or feature requests
Read
CONTRIBUTING.mdto send your first Pull RequestExplore our open collaboration page to join community research efforts