Oumi

What it does

Oumi is a hosted AI model-development platform and open-source toolkit for building custom models around specific tasks. A user can describe a task in natural language, bring a dataset or start without one, and use the platform to evaluate models, inspect failure patterns, synthesize training data, select training settings, fine-tune, and deploy. The workflow produces evaluations and test sets, training data, trained model weights, and deployable models; steps and recipes can be reviewed and modified.

Oumi is aimed at machine-learning teams and domain experts developing models for business-critical or specialized tasks, including teams seeking control over model deployment and iteration. The hosted platform is available through a browser; Oumi also provides a command-line interface and an open-source stack. Deployment options described include Oumi-hosted GPUs and customer infrastructure, including VPC and on-premises for Enterprise. Getting started is through free platform signup; Enterprise inquiries use a contact form.

The Free plan includes up to $50 in credits for three months, while Pro is listed at $25 per month plus pay-as-you-go charges and includes $25 in platform-wide credits per month. Usage prices vary by activity and model: listed meters include agent tokens, evaluation judgments, synthesis rows, storage, training tokens or GPU-hours, and hosted-model inference. Bring-your-own-provider keys shift those providers’ inference charges to the user; without a key, Oumi says it passes through listed provider prices at cost.

What it can do

• Evaluate models and generate test sets and evaluation judges
• Analyze model failures and surface failure modes
• Synthesize and curate targeted training data
• Suggest model family, size, hyperparameters, and algorithms
• Fine-tune models, including SFT and PEFT methods such as LoRA and QLoRA
• Deploy trained models and retrain from production signals

Before you choose

• Free-plan credits are time-limited; the page states up to $50 for three months, and its FAQ gives different credit amounts based on corporate versus personal email.
• Pro includes monthly platform credits but also pay-as-you-go charges; actual usage may incur metered charges for tokens, judgments, rows, storage, training, or inference.
• Using OpenAI, Anthropic, or Google models may require the user’s provider key and incurs provider billing; without a key, Oumi says it passes through the provider’s published price at cost.

Listed under

• Coding
• Sandboxes and infrastructure
• AI
• Data synthesis
• Developers
• Enterprise
• Fine-tuning
• LLM evaluation

Get started

The Free plan offers time-limited credits; continued entitlement after credit expiry is unconfirmed.

1. Create a free account on the Oumi platform; the pricing page lists up to $50 in credits for three months.
2. Describe the model task and target outcome; bring data or begin without it.
3. Review and adjust the generated evaluation, data, and training recipes before running them.

The page lists usage-based charges in addition to plans. SFT is priced per 1M training tokens; on-policy distillation and production inference are priced per GPU-hour. Hosted-model inference is charged when used for platform actions. External OpenAI, Anthropic, or Google usage may be billed directly by the provider when using a user-supplied key.

Source: oumi.ai/pricing, checked 1 Nov 2026. Prices can change; check before you buy.

Sources for this overview

Product pages, documentation, and pricing checked 1 Nov 2026. Each source supports part of the overview; availability and prices can change.

• WELCOME TO OUMI - Oumi ↗docs.oumi.ai
• Oumi – Your AI factory ↗www.oumi.ai
• Pricing – Oumi ↗www.oumi.ai
• Company – Oumi ↗www.oumi.ai
• Contact – Oumi ↗www.oumi.ai
• Oumi – Your AI factory ↗www.oumi.ai

Company website

oumi.ai: product, pricing and policy pages

YC directory mirror

Batch and company name

Social accounts

Accounts the company links, and its posts on X

Human review

Checked by a person

knowledge_base

manual_review

Last checked 1 Nov 2026. Every field keeps its source and date; unknown means we couldn’t confirm it.