Oumi
An end-to-end platform for building foundation models: fine-tuning, evaluating and deploying open-weight LLMs such as Qwen, Gemma and Llama, with reinforcement learning, tool use and MCP integration.
Oumi is a platform for fine-tuning, evaluating and deploying open-weight language models such as Qwen, Gemma and Llama, covering the full path from training to deployment. It supports supervised fine-tuning, direct preference optimization, and GRPO reinforcement learning, including reinforcement learning extended to support tool use.
It provides tools, simulated and lookup environments, and agentic data synthesis for building and training agents, along with an oumi-mcp MCP server for integration with Claude and Cursor, and an oumi deploy command for dedicated inference endpoints on providers such as Fireworks and Parasail. Batch inference is supported across Anthropic, Fireworks and Together, and the project tracks current releases of Transformers, TRL, vLLM and veRL.
It is written in Python and released under the Apache-2.0 license.