What is Oumi?
Oumi is a hosted platform for building custom models and agents and improving them over time. Its centrepiece is FAiE, the Frontier AI Engineer agent: you describe the task in plain language, and it proposes evaluators, measures a baseline, flags bad training rows, suggests a model family, size and training method, deploys an endpoint and groups production failures so they can feed the next round. According to the docs it asks for approval before any step that costs money unless you turn on Auto-approve.
On 4 November 2026 Oumi extended FAiE from models to agents. You define the agent's tools, connect them to simulated or live versions, and evaluate its decisions across tool calls and turns; Bring Your Own Harness lets an agent that already runs elsewhere keep its orchestration and connect only for evaluation and training. GRPO reinforcement learning for agents is in private preview.
Oumi PBC is based in Seattle, raised a $10 million seed round in January 2025, and is built on its own Apache 2.0 library of the same name. The platform's recipes run on that library, which is how Oumi backs its claim that you own the weights, the data and the recipes and can leave with them.
How much does Oumi cost?
| Plan | Price | What you get |
|---|---|---|
| Free | $0 | Up to $50 in credits for 3 months; evaluation, data synthesis and LoRA/QLoRA fine-tuning |
| Pro | $25/month | $25/month in credits, then pay-as-you-go; on-policy distillation, concurrent jobs, autoscaling inference, 1 weight download a month |
| Enterprise | Custom | Your own VPC or on-prem, dedicated GPU capacity with throughput SLAs, RL and custom pipelines |
Free: $0 with up to $50 in credits for 3 months ($25 for a personal email). Pro: $25/month including $25/month in credits, then pay-as-you-go, with 1 model-weight download a month. Enterprise: custom, with your own cloud or on-prem. Usage: FAiE agent $6.25 per 1M input and $31.25 per 1M output tokens; evaluation $1 per 1,000 judgments; data synthesis $1 per 1,000 rows; fine-tuning $0.49 to $6.00 per 1M training tokens by model size; on-policy distillation $2.90 (A100) or $4.00 (H100) per GPU-hour; production inference $3.00 to $10.00 per GPU-hour.
Is Oumi worth it?
Your support bot runs on a frontier model that costs too much per answer and still gets your refund rules wrong. The usual fix, training a smaller model on your own examples, needs people who know how to build evaluations, clean data and pick a training method. Most teams do not have them.
Oumi's agent, FAiE, does that planning for you. You say what the model or agent should do, and it proposes how to score it, measures where it stands, finds the examples it fails on, generates more like them, trains a candidate and compares it against the baseline. Every step waits for your sign-off unless you switch on Auto-approve. Since 4 November it does the same for agents, scoring their tool calls across a whole conversation, and an agent that already runs elsewhere can stay there.
The pitch is cost. Oumi's own case studies claim a 3B model matched Qwen3-235B on SQL at 10 times lower cost for Kaizen Gaming. These are Oumi's numbers, not independent ones. The more useful promise is that the work is portable: the recipes run on Oumi's open-source library, so in principle you can rebuild the model without the platform.
Entry is cheap. Pro is $25 a month with $25 of credits. Fine-tuning a model up to 16B parameters costs $0.49 per million training tokens, so 10 million tokens of examples over three passes comes to about $14.70. The FAiE agent itself is billed at $6.25 per million input tokens and $31.25 per million output, and every evaluation judgment is $1 per 1,000.
The catch: the cheap part is training; the expensive part is keeping a model online. A dedicated H100 for serving is $8.00 an hour, about $5,760 for a month left running, and the price list puts it above the newer H200 at $6.50, which reads like an error. Pro allows one model-weight download a month, so leaving with your weights is real but rationed. On data, the trust centre lists SOC 2 Type 2, SOC 3 and HIPAA reports behind an access request, and names OpenAI, Anthropic, AWS and Google among subprocessors, but Oumi publishes no statement on whether customer data is used to train its own models. GRPO for agents is in private preview, and the homepage still shows a placeholder where one customer's cost saving should be.
Our take: Oumi makes sense when a single narrow task already runs up a large model bill and you can describe what a right answer looks like. For a team that wants to tune prompts and watch traces, a tracing tool such as Braintrust is the smaller step.
Where it earns its place
A product lead at a 40-person insurer paying a frontier model to read claim forms
Get the same extraction quality from a cheaper model we control
Describe the fields to FAiE, upload a few hundred labelled forms, let it build evaluators and a baseline, then fine-tune a small open model. Training under 16B parameters is $0.49 per 1M tokens; serving on a dedicated A100 is $3.00 an hour, so the saving depends on volume.
An engineer at a travel start-up whose booking agent picks the wrong tool
Find out why the agent fails and fix the model, not just the prompt
Connect the existing agent through Bring Your Own Harness, so its orchestration stays put, and let FAiE score tool calls across whole conversations and turn failures into training data. Reinforcement learning with GRPO needs a private-preview request.
A data scientist at a hospital analytics team that cannot send patient data to a public API
Train and run a specialised model inside our own cloud account
Enterprise adds your own VPC or on-prem deployment at a custom price; the trust centre lists a HIPAA report, available on request. Ask for the data-retention policy and contract terms on training before uploading records.
Before you commit
- Setup
- Self-serve sign-up with $50 of credits on a corporate email ($25 personal); describe a task to FAiE and it lays out a plan, proposes evaluators and runs a baseline before anything is trained.
- Integrations
- Open models such as Qwen, Gemma, Llama, DeepSeek, GLM and Kimi hosted by Oumi; OpenAI, Anthropic and Google models through your own keys or at the provider's price; deployed endpoints for your app; agents in other harnesses via Bring Your Own Harness; the open-source Oumi library and its oumi-mcp server.
- Data and privacy
- Trust centre lists SOC 2 Type 2, SOC 3 and HIPAA reports and a data-retention policy, all behind an access request; subprocessors include ClickHouse, Google, OpenAI, Anthropic and AWS. Oumi publishes no statement on whether customer data is used to train its own models. Enterprise can run in your VPC or on-prem. Which AI tools train on your data?
- Cost at scale
- Training is cheap and serving is not: a dedicated H100 endpoint is $8.00 an hour, about $5,760 a month if left running, on top of FAiE agent tokens at $6.25/$31.25 per 1M and evaluations at $1 per 1,000 judgments.
- Limits
- On-policy distillation runs on 8 GPUs subject to availability; Pro allows 1 weight download a month; GRPO for agents is private preview; several pricing FAQs on the site have no answers yet.
What’s good
- FAiE plans evaluations, data fixes and training runs, with approval at each paid step
- Recipes run on the Apache 2.0 Oumi library, so the work is portable
- Fine-tuning from $0.49 per 1M training tokens for models up to 16B
- Bring Your Own Harness evaluates and retrains agents that run elsewhere
- Bring your own OpenAI, Anthropic or Google key and Oumi charges nothing for that inference
What’s not
- Agent building launched on 4 November 2026; GRPO for agents is private preview
- Pro allows only 1 model-weight download a month
- Nine separate meters on the price list make the monthly bill hard to forecast
- No published statement on training on customer data
- Performance claims come from Oumi's own case studies
Best for
Product teams without ML specialists who want to replace an expensive frontier model on one narrow, well-defined task with a smaller model they own.
Oumi alternatives
Oumi: common questions
- Is Oumi worth it?
- Oumi is a credible way for a team without ML engineers to train its own model, but the agent features are less than two weeks old and the bill has nine separate meters. We rate it 3.8 out of 5.
- How much does Oumi cost?
- Free: $0 with up to $50 in credits for 3 months ($25 for a personal email). Pro: $25/month including $25/month in credits, then pay-as-you-go, with 1 model-weight download a month. Enterprise: custom, with your own cloud or on-prem. Usage: FAiE agent $6.25 per 1M input and $31.25 per 1M output tokens; evaluation $1 per 1,000 judgments; data synthesis $1 per 1,000 rows; fine-tuning $0.49 to $6.00 per 1M training tokens by model size; on-policy distillation $2.90 (A100) or $4.00 (H100) per GPU-hour; production inference $3.00 to $10.00 per GPU-hour.
- What are the downsides of Oumi?
- Agent building launched on 4 November 2026; GRPO for agents is private preview; pro allows only 1 model-weight download a month; nine separate meters on the price list make the monthly bill hard to forecast; no published statement on training on customer data; performance claims come from Oumi's own case studies.
- Who is Oumi best for?
- Product teams without ML specialists who want to replace an expensive frontier model on one narrow, well-defined task with a smaller model they own.
- How much does Oumi cost for a team or in production?
- Training is cheap and serving is not: a dedicated H100 endpoint is $8.00 an hour, about $5,760 a month if left running, on top of FAiE agent tokens at $6.25/$31.25 per 1M and evaluations at $1 per 1,000 judgments.
- Is Oumi safe for company data?
- Trust centre lists SOC 2 Type 2, SOC 3 and HIPAA reports and a data-retention policy, all behind an access request; subprocessors include ClickHouse, Google, OpenAI, Anthropic and AWS. Oumi publishes no statement on whether customer data is used to train its own models. Enterprise can run in your VPC or on-prem.
- What are the best Oumi alternatives?
- The closest alternatives we have reviewed are Together AI, Fireworks AI and Braintrust.
Answers checked Nov 16, 2026.