A Seattle startup founded by former Google and Microsoft engineers is betting that companies will increasingly want to build and own smaller, specialized AI models rather than relying on general-purpose alternatives from the likes of OpenAI, Anthropic, and Google.
Oumi, which launched its open-source platform last year, moved into its commercial phase Tuesday with technology that it says can automate the process of building custom AI models in hours rather than the weeks or months it would take with traditional approaches.
The company says users can describe what they want a model to do in plain language, and the platform handles everything from generating training data to fine-tuning and evaluation.
“One of the last things still standing that has not been automated with AI is the building of AI itself,” said Oumi CEO Manos Koukoumidis, demonstrating the platform on a call this week.
The case for custom models: The argument comes down to cost, performance, and control. General-purpose models from companies like OpenAI, Anthropic, and Google are designed to handle a wide range of tasks but can fall short on specialized work.
Smaller models trained on specific tasks can be cheaper to run, faster to respond, and deployed on a company’s own infrastructure, keeping sensitive data in-house.
Company background: As first reported by GeekWire last year, Oumi was launched by Koukoumidis, a former senior engineering manager at Google Cloud, and Oussama Elachqar, a machine learning engineer who previously worked at Apple, Twitter, and Microsoft.
Other co-founders are Matthew Persons and Jeremiah Greer, both former Google AI and Microsoft engineers; William Zeng, who worked on AI safety at Google; and Kostas Aisopos, a former Google staff engineer who also previously worked at Microsoft.
The company raised $10 million last year in a seed round led by Venrock and Obvious Ventures, with participation from Plug & Play and Seattle-based Ascend.
The company has since grown to about 20 employees, based primarily in Seattle.
Broader landscape: It’s part of a growing market for tools that help companies build and deploy custom AI models. Predibase, a model fine-tuning platform, was acquired by Rubrik last year for more than $100 million. Fireworks AI offers inference and fine-tuning for open-source models.
Open-source projects like Unsloth and Axolotl provide alternatives geared for developers.
Major cloud providers including Amazon and Google offer similar capabilities through technologies such as AWS SageMaker and Google’s AI Studio, although they often require more setup and tend to tie customers to those specific cloud ecosystems.
Part of Oumi’s pitch is that its natural-language interface and end-to-end automation make the process accessible to many other users beyond dedicated AI engineers.
The company launched as an open-source project in early 2025, and its toolkit has since drawn nearly 9,000 GitHub stars and adoption at research institutions including Stanford, MIT, and Berkeley.
Commercial expansion: The new launch builds on that open-source foundation.
Oumi says smaller, custom-trained models can outperform the largest general-purpose models on specific tasks, at a fraction of the cost. It points to early case studies to show the potential.
- Divisions Maintenance Group, a facility services company, used the platform to fine-tune a model with fewer than 1 billion parameters for invoice validation. Accuracy jumped from 72% to 99%, matching the performance of OpenAI’s GPT-5.2 on the task.
- Aurasell, an AI-powered CRM platform, built an 8 billion parameter model for extracting information from web pages that outperformed Anthropic’s Sonnet 4.5 on key metrics.
The Oumi Platform starts at $25 per month per organization with unlimited users, plus pay-per-use fees for training, evaluation, and data synthesis.
