Oumi

Oumi helps ML researchers and teams build specialized AI models for their own tasks. It brings data preparation, training, evaluation, and deployment into one platform, with reusable recipes and components you can customize.

• Model training: Fine-tune text and vision-language models such as Qwen, Llama, Gemma, and DeepSeek using LoRA, QLoRA, or reinforcement learning, with distributed training for larger workloads.
• Training data: Generate and curate examples with LLM judges, including multi-turn conversations for models that call tools.
• Evaluation: Compare models on standard benchmarks or custom criteria, and inspect the results, training examples, and recipes.
• Serving and integrations: Serve models with vLLM or SGLang, connect to OpenAI, Anthropic, and Vertex AI APIs, or use the MCP server with Claude or Cursor.
• Automated model development: The hosted service works from a plain-English task description to automate evaluation, data generation, training choices, and deployment, then uses production feedback to support retraining on a schedule you control.

The open-source stack uses the Apache 2.0 license and runs on laptops, clusters, and cloud infrastructure. You own the resulting model weights, data, and recipes. Deployment can use Oumi's cloud GPUs or your own servers, including a private cloud or on-premises infrastructure; specialized models can also run on smartphones and embedded devices. Commercial API integrations send requests to the selected external provider.