# oumi-ai/oumi

Easily fine-tune, evaluate and deploy Qwen, Gemma, or any open weight LLM!

Repository: https://github.com/oumi-ai/oumi
Canonical: https://ross.abutalabs.com/products/oumi
Homepage: https://oumi.ai
Language: Python
License: Apache-2.0
License Family: permissive
Topics: dpo, evaluation, fine-tuning, inference, llama, llms, sft, vlms, gpt-oss, gpt-oss-120b, gpt-oss-20b, slms, open-weight, open-weight-models, open-weights
Last push: 2026-08-27T00:02:41+00:00

## Health v2 (maintenance only)
Score: 86/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 99, release rhythm 83, longevity 60
- inputs: {"age_days": 848, "days_push": 7, "days_rel": 118, "gap_med": 13.0, "n_releases_24m": 23}
- flags: none
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 9373, forks 784 (observed 2026-08-28T04:10:30.924702+00:00)

## What it is
Oumi is an open-source Python framework and platform for the end-to-end lifecycle of open-weight LLMs: data synthesis, fine-tuning (SFT, LoRA, DPO, GRPO), evaluation, and deployment. It supports models like Qwen, Gemma, Llama, and GPT-OSS, with a CLI, an MCP server, and integrations with vLLM, TRL, and cloud GPU providers.

## Use cases
- fine-tune llama on my own dataset
- run sft or dpo training on an open weight llm
- evaluate an llm on a custom evaluation dataset
- deploy a fine-tuned model as an inference endpoint
- synthesize training data with an llm
- train a small language model for a specific task
- distill a frontier model into a smaller one
- run grpo reinforcement learning with tool use

## When to choose
- you want one open-source tool covering data, training, evaluation, and deployment for open-weight LLMs
- you need to fine-tune Qwen, Gemma, Llama, or VLMs with SFT, LoRA, DPO, or GRPO
- you want to deploy dedicated inference endpoints or run inference with vLLM
- you prefer Apache-2.0 licensing and Python-based configuration

## When to avoid
- you only need prompt-based access to closed frontier models via an API
- you need a no-code GUI-only training experience
- you lack GPU resources and don't want to use cloud training
- you need a lightweight inference-only server without training features

## Facets
- artifact type: framework
- maturity: active
- function: llm-training, llm-inference, machine-learning, rag, mcp, cli, data-generation
- domain: large-language-models, machine-learning, deep-learning, artificial-intelligence, developer-tools
- platform: python, cli, cloud
- tags: fine-tuning, sft, dpo, grpo, reinforcement-learning, open-weights, vlms, model-deployment, distillation, vllm, peft, lora, evaluation, gpu, docker

## Member repositories
- oumi-ai/oumi (main) score 86

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:10:30.924702+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-29T17:22:21.557331+00:00, confidence not recorded.
  - readme: https://github.com/oumi-ai/oumi (fetched 2026-08-28T04:10:30.924702+00:00, sha 08c852d2527c)
  - homepage: https://oumi.ai (fetched 2026-08-29T08:22:29.643609+00:00, sha c877bdea1b42)
  - site_page: https://docs.oumi.ai (fetched 2026-08-29T08:22:29.648757+00:00, sha 4a63c8db6db2)
  - registry_pypi: https://pypi.org/pypi/oumi/json (fetched 2026-08-29T08:22:29.652889+00:00, sha 56afb6917423)
  - site_page: https://oumi.ai/pricing (fetched 2026-08-29T08:22:29.646802+00:00, sha ef4b405cd790)
  - site_page: https://oumi.ai/company (fetched 2026-08-29T08:22:29.650976+00:00, sha af081c7614a7)
- Data as of 2026-08-30T08:39:29.467469+00:00.