# Oumi review

> Oumi is an open-source Python platform for fine-tuning, evaluating, and deploying open text and multimodal models, including agentic tool-using LLMs.

Devtoolwatch Board, the week of Oct 25–31, 2026. Canonical page: https://index.devtoolwatch.com/tools/oumi-ai-oumi/

## TL;DR

Oumi is an Apache-2.0 platform covering the model lifecycle: data preparation, training, evaluation, inference, and deployment. It is for ML engineers and researchers who fine-tune open models such as Llama, Gemma, and Qwen. It sets itself apart with ready-made YAML recipes, one CLI that runs from laptops to clouds, and tool-use training with GRPO over executable environments.

## Oumi at a glance

- **Category:** [Agent frameworks](https://index.devtoolwatch.com/categories/agent-frameworks/)
- **Maintained by:** Oumi (organization, [@oumi-ai](https://github.com/oumi-ai) on GitHub)
- **License:** Apache 2.0
- **Language:** Python, Jupyter Notebook, Shell
- **Repository:** [github.com/oumi-ai/oumi](https://github.com/oumi-ai/oumi)
- **Website:** [oumi.ai](https://oumi.ai)
- **PyPI package:** [oumi](https://pypi.org/project/oumi/) 0.9
- **Latest release:** [v0.9](https://github.com/oumi-ai/oumi/releases/tag/v0.9) on GitHub, Sep 16, 2026
- **First published:** May 7, 2024

## What Oumi does

- Oumi trains and fine-tunes models from 10M to 400B+ parameters using SFT, LoRA, QLoRA, DPO, GRPO, and other techniques.
- Oumi trains agentic models with tool-calling SFT and runs GRPO with verl over database, HTTP, lookup, and simulated tool environments.
- Oumi synthesizes and curates training data, including multi-turn tool-use conversations, using LLM judges and the `oumi synth` module.
- Oumi evaluates models on standard benchmarks and with multi-criteria rubric judges such as `RubricJudge`.
- Oumi serves models with vLLM or SGLang and deploys them to dedicated endpoints on Fireworks and Parasail via `oumi deploy`.
- Oumi launches remote jobs on AWS, Azure, GCP, Lambda, Modal, Slurm, and other platforms with `oumi launch`.
- Oumi provides an `oumi-mcp` MCP server so Claude or Cursor can drive Oumi workflows.

### How it works

Oumi runs as a Python package and CLI. You point commands such as `oumi train`, `oumi evaluate`, and `oumi infer` at YAML config recipes. Jobs run locally, in Docker, or on cloud clusters through `oumi launch`. Oumi connects to open models and commercial APIs such as OpenAI, Anthropic, Vertex AI, Together, Fireworks, and Parasail. It supports FSDP, DeepSpeed, and DDP for distributed training.

## How to install Oumi

**With uv pip**

```
uv pip install oumi
```

**With GPU support**

```
uv pip install 'oumi[gpu]'
```

**With Docker**

```
docker pull ghcr.io/oumi-ai/oumi:latest
```

### Requirements

- Python 3.10 or later and below 3.15.
- A GPU for the `oumi[gpu]` install and the Docker commands, which use `--gpus all`.
- API keys for hosted engines, such as the FIREWORKS_API_KEY environment variable for Fireworks.
- A cloud account (AWS, Azure, GCP, Lambda, and others) for remote jobs via `oumi launch`.

### First steps

1. Train a small model with the quickstart recipe: oumi train -c configs/recipes/smollm/sft/135m/quickstart_train.yaml
2. Evaluate it with: oumi evaluate -c configs/recipes/smollm/evaluation/135m/quickstart_eval.yaml
3. Chat with a model interactively: oumi infer -c configs/recipes/smollm/inference/135m_infer.yaml --interactive
4. Open the Getting Started tour notebook in Colab to explore training, evaluation, inference, and job management.

## Who Oumi is for

- ML engineers fine-tuning open models like Llama, Gemma, Qwen, or DeepSeek from config recipes.
- Teams training agentic, tool-using models with tool-calling SFT and reinforcement learning.
- Researchers who need reproducible experiments that scale from a laptop to cloud clusters.
- Developers distilling large models into smaller, specialized ones for deployment.

### Limitations

- Oumi focuses on training and serving models, not on building agent applications at runtime.

## Usage data (the week of Oct 25–31, 2026)

- Board rank: #304 overall, #30 in Agent frameworks, momentum score 31
- Downloads that week: 9.6k (+28% vs 4-week average)
- New GitHub stars per week (last 12 weeks): Aug 9: 11; Aug 16: 3; Aug 23: 9; Aug 30: 4; Sep 6: 4; Sep 13: 5; Sep 20: 4; Sep 27: 2; Oct 4: 9; Oct 11: 12; Oct 18: 14; Oct 25: 12

## Oumi alternatives

| Rank | Tool | Stars | Downloads/wk |
|---|---|---|---|
| #1 | [Harness²](https://index.devtoolwatch.com/tools/google-research-harness2/) | 21.7k | – |
| #2 | [HyperFrames](https://index.devtoolwatch.com/tools/heygen-com-hyperframes/) | 76.6k | 1.5M |
| #3 | [AgentScope](https://index.devtoolwatch.com/tools/agentscope-ai-agentscope/) | 35k | 51.4k |
| #4 | [AG-UI](https://index.devtoolwatch.com/tools/ag-ui/) | 16.9k | 4.6M |
| #5 | [LangGraph](https://index.devtoolwatch.com/tools/langgraph/) | 44.2k | 18.5M |
| #6 | [Bifrost](https://index.devtoolwatch.com/tools/maximhq-bifrost/) | 9.3k | – |

## Oumi FAQ

### Is Oumi free and open source?

Yes. Oumi's source code is on GitHub (oumi-ai/oumi) under the Apache 2.0 license, so you can use it for free. Its website also links to a pricing page (https://oumi.ai/pricing) for any paid plans.

### How do I install Oumi?

Run `uv pip install oumi`. The oumi package needs Python 3.10 or newer, below 3.15. The install section above has the other options from the project's README.

### Is Oumi still maintained?

Yes. The oumi-ai/oumi repository was last updated on Oct 21, 2026, and the latest release, v0.9 on GitHub, came out on Sep 16, 2026.

### How popular is Oumi?

Oumi has 9.4k GitHub stars as of Nov 1, 2026. It gained 12 new stars in the week of Oct 25–31, 2026, against a 12-week average of 7 a week. Its main package had 9.6k downloads that week, +28% against its own 4-week average. That made it #30 among AI agent frameworks on the Devtoolwatch Board.

### What are the best alternatives to Oumi?

Other AI agent frameworks on the Devtoolwatch Board in the week of Oct 25–31, 2026, by momentum: Harness² (#1), HyperFrames (#2), AgentScope (#3), AG-UI (#4), LangGraph (#5). The alternatives table above has their numbers.

### How does Oumi work?

Oumi works through a CLI and Python API driven by YAML configs. You run `oumi train`, `oumi evaluate`, `oumi infer`, or `oumi launch` with a recipe. It handles data synthesis, training, evaluation, inference, and deployment for open text and vision-language models.

### What is the difference between Oumi and HallOumi?

Oumi is the open-source platform for building and training models. HallOumi is a claim-verification model that the Oumi team introduced in April 2025.

### Can Oumi train tool-using agent models?

Yes. Since v0.9, Oumi supports tool-calling SFT and DPO, executable tool environments (database, HTTP, lookup, simulated), multi-turn tool conversation synthesis, and GRPO over those environments with verl.

### Can I use Oumi from Claude or Cursor?

Yes. Oumi v0.8 added the `oumi-mcp` MCP server, which lets MCP-capable assistants such as Claude and Cursor drive Oumi.

### Which cloud platforms can Oumi run jobs on?

Oumi runs remote jobs on AWS, Azure, GCP, Lambda, Modal, Slurm, and others through `oumi launch`. Version 0.4.2 also added the NERSC Perlmutter HPC cluster.

## Sources

1. [oumi-ai/oumi on GitHub](https://github.com/oumi-ai/oumi)
2. [Oumi README](https://github.com/oumi-ai/oumi/blob/main/README.md)
3. [oumi on PyPI](https://pypi.org/project/oumi/)
4. [Oumi website](https://oumi.ai)
5. [Release v0.9 (Sep 16, 2026)](https://github.com/oumi-ai/oumi/releases/tag/v0.9)
6. [Release v0.8 (May 7, 2026)](https://github.com/oumi-ai/oumi/releases/tag/v0.8)
7. [Release v0.7 (Jan 29, 2026)](https://github.com/oumi-ai/oumi/releases/tag/v0.7)
8. [Hacker News: Oumi-AI/oumi: Everything you need to build foundation models, end-to-end](https://news.ycombinator.com/item?id=42879844)
9. [Hacker News: Everything you need to build foundation models, end-to-end](https://news.ycombinator.com/item?id=42909538)
10. [Hacker News: Small Fine-Tuned Models Are All You Need](https://news.ycombinator.com/item?id=45637668)
11. [Devtoolwatch Board, the week of Oct 25–31, 2026](https://index.devtoolwatch.com/)