oumi-ai/oumi

Oumi – Open‑source platform for the whole lifecycle of foundation models

What it is – Oumi is a Python‑based, fully open‑source toolkit that lets you prepare data, train, fine‑tune, evaluate, and deploy large language models (LLMs) and vision‑language models. It targets everything from tiny 10 M‑parameter experiments on a laptop to 400 B‑parameter runs on a GPU cluster, and it supports both text‑only and multimodal models.

Core capabilities

• Training & fine‑tuning – SFT, LoRA, QLoRA, DPO, GRPO, and other state‑of‑the‑art methods. Distributed training via FSDP, DeepSpeed, DDP, and support for PyTorch 2.9+.
• Agentic / tool‑using models – Built‑in tool‑calling SFT, executable tool environments (database, HTTP, lookup, simulated), and reinforcement learning with the verl GRPO algorithm.
• Multimodal support – Works with Llama, Gemma, Qwen, DeepSeek, GLM, Phi, and many vision‑language families (Qwen‑VL, InternVL, etc.).
• Data synthesis & judging – LLM‑as‑a‑judge pipelines, multi‑turn tool‑use conversation generation, and a rubric‑based RubricJudge for quality control.
• Inference & deployment – Fast local inference via vLLM or SGLang, plus a oumi deploy CLI that can spin up dedicated endpoints on cloud providers (Fireworks, Parasail, Lambda, Modal, etc.).
• Remote job orchestration – oumi launch abstracts away AWS, Azure, GCP, Lambda, and other clouds, handling resource provisioning and monitoring.
• Extensive recipe library – Ready‑to‑run YAML configs for dozens of models (Qwen 3, Gemma 4, Llama 4, Falcon‑H1/E, etc.) covering training, inference, and evaluation.

How to get started

# Install the core package (CPU only)
uv pip install oumi
# GPU‑enabled install
uv pip install 'oumi[gpu]'
# Or pull the official Docker image
docker pull ghcr.io/oumi-ai/oumi:latest

After installation you can run the CLI:

# Train a 135M model (quickstart config)
oumi train -c configs/recipes/smollm/sft/135m/quickstart_train.yaml
# Evaluate the same model
oumi evaluate -c configs/recipes/smollm/evaluation/135m/quickstart_eval.yaml
# Interactive inference
oumi infer -c configs/recipes/smollm/inference/135m_infer.yaml --interactive

For cloud runs replace train with launch up and add --resources.cloud aws|azure|gcp|lambda.

Why you might choose Oumi

• Zero boilerplate – recipes give you a working pipeline in minutes.
• Enterprise‑grade reliability – used by teams that train models at scale; includes robust logging, checkpoint handling, and fault‑tolerant job management.
• Research flexibility – every stage (data, model, reward, evaluation) is modular, so you can swap components without rewriting code.
• Broad ecosystem – integrates with open‑source models, commercial APIs (OpenAI, Anthropic, Together, etc.), and a variety of inference back‑ends.
• Active community – Apache 2.0 licensed, with Discord, blog, and frequent releases (v0.9 as of Sep 2026).

Documentation & learning resources

• Full docs: https://oumi.ai/docs
• Quick‑start notebooks (Colab links) for training, distillation, evaluation, vision‑language fine‑tuning, and GRPO RL.
• Blog posts and webinars covering use‑cases like text‑to‑SQL fine‑tuning and claim‑verification models.

License – Apache 2.0 (permissive, commercial‑friendly).

Status – Actively maintained; beta label indicates rapid development, but core features (training, evaluation, inference) are stable.

---

All information above is taken directly from the repository’s README.