Oumi

Oumi is an open-source platform that covers the full lifecycle of foundation models: fine-tuning, evaluation, data synthesis and deployment. Developed by Oumi AI and hosted on GitHub with more than 9,400 stars, it aims to make training open models as routine as training any other software, and it is updated daily through 2026.

Overview

Instead of stitching together separate scripts for every stage, Oumi provides one consistent interface driven by configuration files and a friendly CLI. You can fine-tune models such as Llama, Qwen and Gemma with SFT, LoRA or preference methods like DPO and GRPO, evaluate them with standard harnesses and LLM-as-a-judge setups, generate synthetic training data, and serve the result for inference. The same configuration runs on a laptop for a smoke test and scales to multi-node GPU clusters on major clouds, which makes experiments reproducible from first prototype to production run.

Key Features

• Unified workflow for training, evaluation, synthesis and inference of LLMs and vision-language models
• Modern training recipes: SFT, LoRA/QLoRA, DPO, GRPO and other reinforcement learning methods
• Support for popular open model families including Llama, Qwen, Gemma and gpt-oss
• Built-in evaluation with standard benchmarks plus configurable LLM-as-a-judge scoring
• Scales from a single workstation to distributed clusters across cloud providers
• Configuration-driven design that keeps experiments reproducible and shareable

💳 Pricing

Oumi is free and open source under the Apache License 2.0. There is no paid edition; costs are limited to the compute you choose to run it on.

Who Is It For

Oumi is for ML engineers and research teams who fine-tune and evaluate open models regularly and want one well-maintained platform instead of a drawer full of one-off scripts. It also suits practitioners moving from API-only usage to owning their models, since it smooths the whole path from data to deployment.