Oumi
Oumi is a Seattle startup whose open-source platform and hosted service let a product team fine-tune, evaluate and serve its own small models in place of frontier APIs.
Founder Stack
Listed in AI
- Contact
Is this your organization? Claim it
Facts
- HQ
- Seattle
- Sectors
- Software Development
- Founded
- 2024
- Employees
- 2–10
- Registered name
- OUMI PBC
- Stage
- Seed
- Founders
- Manos Koukoumidis, Matthew Persons, Jeremiah Greer, William Zeng, Kostas Aisopos
Oumi Adds Agent Development, BYOH, Advanced Training, and Cost Visibility in November Update
Oumi’s November product updates add agent development, Bring Your Own Harness, advanced training, and cost visibility to its platform, expanding tools for building and managing AI agents.
Marketer Builds Weekend Claims-Triage AI on Oumi That Routes 92% of Test Claims, Beating Claude Sonnet's 78%
A marketer with limited ML experience used Oumi over a weekend to build a claims triage model that correctly routed 92% of test claims, outperforming Claude Sonnet which achieved 78%.
Raffi Krikorian (Mozilla) and Manos Koukoumidis (Oumi) Argue Building AI Ownership Beats Renting for Long-Term Value
Raffi Krikorian and Manos Koukoumidis debate whether organizations should rent AI services or build their own, arguing that long-term ownership of intelligence captures greater value.
TechCrunch Disrupt 2026 Recap: Every Company Will Build Its Own AI — What to Build and How to Do It
The author argues that every company will develop its own AI, outlining what to build and how, and shares founders' questions from TechCrunch Disrupt 2026 to illustrate practical concerns and implementation approaches.
Analysis: TypeSafe’s Jev Likely Delivers One-Token, Calibrated Decisions and Runs on Qwen 3.5
The post analyzes how TypeSafe’s Jev decision model probably operates, arguing it produces one-token answers with calibrated probabilities, serves results quickly, and is likely built on Qwen 3.5.
Oumi’s Ryan Arman: Fine-tuned Qwen3 4B Hits ~93% on BANKING77, Jev Reaches 92.4% Only with Retrieval of 24 Examples
The authors compare Jev (a retrieval-augmented decision-making approach) with a small fine-tuned Qwen3 4B model on the BANKING77 intent benchmark and find that a 30-minute fine-tune reached ~93% accuracy while Jev needed retrieval of many labeled examples to approach 92.4%, highlighting a tradeoff between avoiding training and keeping systems simple.
Oumi Announces September 2026 Product Updates Including CLI and New Model
Oumi provides product updates for September 2026, highlighting a CLI and a new model as key enhancements.
Oumi launches BYOC with Tensor9 so enterprises can run full Oumi Platform inside their VPC, keeping data local
Erik Bower and Min Song announce that Oumi’s full AI platform can now run inside customers’ own cloud or on-premises (BYOC) using Tensor9, keeping data stored and processed entirely within the customer VPC under their IAM, region, and controls.
Oumi Announces New Models in August 2026, Directing Readers to Full Post on oumi.ai
Oumi posts brief August 2026 updates about new models, directing readers to read the full post on oumi.ai.
Oumi announces August 2026 updates for its AI that improves with every run
The short post announces August 2026 updates about an AI product called Oumi, indicating the system improves with each run and points readers to the full post on oumi.ai.
People
See all 6-
Manos Koukoumidis Co-founder, Cofounder - CEO
-
Oussama Elachqar Co-Founder
-
Jeremiah Greer Co-founder, Cofounder - Head of Research