What is Oumi?
Oumi is an open-source Python library that covers the whole post-training loop for open models: synthesising and curating data with LLM judges, training, evaluating against benchmarks or rubric judges, and running inference. One YAML recipe drives each step through the `oumi train`, `oumi evaluate`, `oumi infer` and `oumi launch` commands, and the README claims support for models from 10M to 400B+ parameters, text and multimodal, including Llama, Gemma, Qwen, DeepSeek, GLM and Phi.
Training methods include SFT, LoRA, QLoRA, DPO, KTO and GRPO, with FSDP, DeepSpeed and DDP for scale. Version 0.9, released on 16 September 2026, turned it towards agents: tool-calling SFT and DPO, executable tool environments backed by databases, HTTP endpoints, lookup tables or a simulating model, GRPO over those environments through verl, and a multi-criteria RubricJudge.
Inference runs on vLLM or SGLang locally, or on OpenAI, Anthropic, Vertex AI, Together, Fireworks and others remotely; `oumi deploy` creates dedicated endpoints on Fireworks and Parasail, and an `oumi-mcp` server lets Claude or Cursor drive it. Jobs launch on AWS, Azure, GCP, Lambda, Modal or Slurm. It needs Python 3.10 to 3.14 and is maintained by Oumi PBC, whose paid platform runs the same recipes.
Is Oumi worth it?
Oumi's case is breadth. Where Axolotl and Unsloth concentrate on the training run, Oumi also generates and filters the data, scores the result with LLM judges and serves the model, all from the same YAML recipe format and one CLI. That makes it the natural choice when the bottleneck is the loop around training rather than the training itself.
Version 0.9, released on 16 September 2026, is the reason to look now. It adds tool-calling SFT and DPO, environments in which tool calls actually execute against a database session (rolled back after each episode), an HTTP endpoint or a deterministic lookup table, and GRPO over any of those environments through verl. Few open libraries connect agent data synthesis, an executable environment and reinforcement learning in one place.
Adoption is moderate. At 9,437 GitHub stars it trails Axolotl (12,728) and Hugging Face TRL (19,744), on which it partly builds, and the README still labels the project beta, warning that advanced features may change. Releases arrive every four months or so: v0.7 in January, v0.8 in May and v0.9 in September 2026. Open issues and pull requests together number 26, a small backlog for a project this wide.
Two defaults deserve attention. Anonymous usage analytics are on unless you set DO_NOT_TRACK=1 or edit ~/.oumi/telemetry.json; the docs say they cover commands, hardware, package versions and stack traces, never training data, weights or file paths. And the library is the open core of a commercial platform, so the FAiE agent that plans runs for you, managed GPUs and hosted endpoints sit on the paid side.
Where it earns its place
An ML engineer at a 25-person SaaS company with a ticket-routing agent
Teach a small open model to call our internal tools correctly
Synthesise multi-turn tool conversations, run tool-calling SFT, then GRPO against an EndpointEnvironment that hits a staging API, scoring with a RubricJudge. The library is free; the cost is GPU hours on whichever cloud `oumi launch` targets.
A university lab benchmarking vision-language models
Fine-tune and evaluate several open VLMs with the same recipe format
Start from the bundled recipes for Qwen, Gemma or InternVL, run `oumi train` and `oumi evaluate` on a Slurm cluster, and keep each YAML beside the results for reproducibility. Set DO_NOT_TRACK=1 on shared machines.
A data team at an e-commerce company cleaning product-description training data
Filter thousands of noisy examples before fine-tuning
Use the built-in LLM judges, on a local vLLM model or a commercial API key, to score and drop weak rows, and `infer_partial()` to keep partial results if a long batch fails halfway.
Before you commit
- Setup
- `uv pip install oumi` (or `oumi[gpu]`), the ghcr.io/oumi-ai/oumi Docker image, or an experimental install script; the quickstart trains a 135M SmolLM recipe with one `oumi train` command, and Colab notebooks cover LoRA, distillation, GRPO and judges.
- Integrations
- Hugging Face models and datasets; vLLM, SGLang and remote engines for OpenAI, Anthropic, Vertex AI, Together, Fireworks, Parasail and OpenRouter; verl for RL; `oumi launch` for AWS, Azure, GCP, Lambda, Modal and Slurm; `oumi deploy` to Fireworks and Parasail; `oumi-mcp` for Claude and Cursor.
- Data and privacy
- Runs on your hardware or cloud account. Anonymous telemetry (commands, hardware, versions, stack traces) is on by default; DO_NOT_TRACK=1 turns it off. Data sent to remote inference engines follows that provider's terms. Apache 2.0 permits commercial use; base-model licences still apply.
- Cost at scale
- The library is free; GPU time is the bill. Oumi publishes no VRAM table, so size jobs from a trial run; the company's hosted platform is the paid route if you want managed GPUs.
- Limits
- Beta status; Python 3.10 to 3.14 and PyTorch below 2.13; very large models need multi-node FSDP or DeepSpeed setups you configure yourself.
What’s good
- Data synthesis, training, judging, evaluation and serving in one recipe format
- Tool-calling SFT, executable environments and verl GRPO since v0.9
- Launches on AWS, Azure, GCP, Lambda, Modal and Slurm
- Apache 2.0, with an oumi-mcp server for Claude and Cursor
What’s not
- Labelled beta; advanced APIs may change
- Anonymous telemetry on by default
- Fewer stars and community recipes than Axolotl or TRL
- Breadth means many dependencies, pinned to PyTorch below 2.13
Best for
ML engineers who want one open toolkit for synthetic data, fine-tuning, judging and serving, especially for tool-using models.
Oumi alternatives
Oumi: common questions
- Is Oumi worth it?
- Oumi covers the whole loop around a fine-tune, from synthetic data to a served endpoint, and since v0.9 it can train tool-using agents end to end; it is still labelled beta. We rate it 4.0 out of 5.
- How much does Oumi cost?
- Free and open source under the Apache 2.0 licence. Oumi's hosted platform, which runs the same recipes, is a separate paid product.
- What are the downsides of Oumi?
- Labelled beta; advanced APIs may change; anonymous telemetry on by default; fewer stars and community recipes than Axolotl or TRL; breadth means many dependencies, pinned to PyTorch below 2.13.
- Who is Oumi best for?
- ML engineers who want one open toolkit for synthetic data, fine-tuning, judging and serving, especially for tool-using models.
- How much does Oumi cost for a team or in production?
- The library is free; GPU time is the bill. Oumi publishes no VRAM table, so size jobs from a trial run; the company's hosted platform is the paid route if you want managed GPUs.
- Is Oumi safe for company data?
- Runs on your hardware or cloud account. Anonymous telemetry (commands, hardware, versions, stack traces) is on by default; DO_NOT_TRACK=1 turns it off. Data sent to remote inference engines follows that provider's terms. Apache 2.0 permits commercial use; base-model licences still apply.
- What are the best Oumi alternatives?
- The closest alternatives we have reviewed are Axolotl, TRL, Unsloth and LlamaFactory.
Answers checked Nov 16, 2026.