Oumi

Oumi, made by Oumi PBC, is an open-source stack and hosted platform for building specialized AI models. The open-source repository is licensed under Apache 2.0 and covers training, fine-tuning, evaluation, inference, and data synthesis driven by YAML configs. The company's platform adds an agent-guided workflow from a plain-English task description to a deployed model.

What It Is

Oumi OSS is aimed at ML engineers and researchers who want one interface for the model lifecycle instead of stitching together separate libraries. It supports training and fine-tuning from small models to 400B+ parameters using SFT, LoRA, QLoRA, DPO, GRPO and other methods. It works with text and multimodal models, including Llama, Gemma, Qwen, DeepSeek, GLM, Phi and gpt-oss families. The README notes the project is in beta, with stable core features.

Workflow and Outputs

The workflow runs from a prompt through evaluation, data synthesis, training, deployment and ongoing retraining. Evaluation covers standard benchmarks, LLM judges and multi-criteria rubric judges. Data synthesis and curation can use LLM judges and include multi-turn tool-use conversations. Inference runs on engines such as vLLM, SGLang and llama.cpp or through commercial APIs, and oumi deploy targets dedicated inference endpoints. Jobs can be launched on AWS, GCP, Azure, Lambda and other clouds with oumi launch.

Agentic Model Training

The v0.9 release, per the README, is an end-to-end stack for agentic, tool-using models. It adds tool-calling SFT/DPO, executable tool environments (database, HTTP endpoint, lookup, simulated), and GRPO over Oumi environments with verl. An oumi-mcp MCP server lets Claude or Cursor drive Oumi.

Platform and Ownership Claims

Oumi's website says users own the resulting weights, data and recipes and can run them on Oumi's GPUs or in their own infrastructure. The company states that specialized models can cost up to 90% less than frontier models and reach up to 50% higher accuracy; these are vendor claims. It also lists a continual-improvement loop in which production failures feed retraining, and says deployments scale GPUs to zero when idle.