Oumi AI

Oumi Unveils First Unconditionally Open AI Platform to Ensure AI is Developed Collectively and Responsibly

January 29th, 2025

Oumi unites AI researchers and developers globally to collaborate, experiment, and build together the world's largest state-of-the-art open-source laboratory.

Seattle, WA - Oumi, an AI laboratory created in collaboration with researchers from 13 leading research universities, including Carnegie Mellon University, Stanford University, and the Massachusetts Institute of Technology, has unveiled the world's first unconditionally open source AI platform. By enabling open collaboration, Oumi enables thousands of researchers, developers, and AI experts to come together for the first time to advance AI collectively and responsibly.

The development of multimodal foundation models has largely been limited to incumbent AI platform companies, which release models as closed, proprietary black boxes, or masquerade open-weight models as "open source." Oumi is the first to launch a platform that offers state-of-the-art foundation models with open code, open data, open weights—and open collaboration—uniting a global community of developers and researchers to drive AI forward responsibly and transparently. Oumi is a Public Benefit Corporation and announced $10 million in Seed funding led by Venrock and Obvious Ventures with contributions from Plug & Play and Ascend.

"AI needs its Linux," said Manos Koukoumidis, CEO and Co-founder of Oumi and previously a Senior Engineering Manager at Google Cloud AI who bootstrapped and led the efforts for Cloud PaLM. "The greatest technology innovations are developed in the open. We must open AI up to the open source community to advance it collectively and responsibly. To make it accessible to everyone."

Oumi is designed to support all common foundation model workflows in one unified platform to promote accessibility and to foster open collaboration with all AI researchers and developers. With the all-in-one Oumi platform, developers can:

  • Train and fine-tune models from 10M to 405B parameters using state-of-the-art techniques (SFT, LoRA, QLoRA, DPO, and more)
  • Work with both text and multimodal models (Llama, Qwen, Phi, and others)
  • Synthesize and curate training data with LLM judges, then readily use it to train models
  • Deploy models efficiently with popular inference engines (vLLM, SGLang)
  • Evaluate models comprehensively across standard benchmarks
  • Run anywhere - from laptops to clusters to clouds (AWS, Azure, GCP, Lambda, and more)
  • Integrate with both open models and commercial APIs (OpenAI, Anthropic, Vertex AI, Parasail, etc.)

The current AI landscape creates the illusion of a talent desert that is now starting to impact innovation, as resources focus on individual, rather than collective, AI projects.

"Our vision with Oumi is to free the AI talent locked in silos and make AI the ultimate team sport," said Oussama Elachqar, Co-founder of Oumi and previously a Machine Learning Engineer at Apple. "Giving AI talent a platform where they can work collectively will accelerate progress and speed of discovery in every area of AI. We have incredibly talented and bright developers and researchers in AI working in silos. If we unite them, we will get further, faster."

"There hasn't been enough focus on open experimentation to collectively push the boundaries of what AI can do," said Ganesh Srinivasan, Partner at Venrock. "With an unconditionally open source platform, we won't cap our talent, we will foster experimentation, and society will reap the benefits."

Oumi is designed to be fully flexible and easy to use. The end-to-end, all-in-one platform supports the full AI lifecycle with one consistent interface - from pretraining to data curation, data synthesis, fine-tuning (SFT, LoRA, QLoRA, DPO), inference, and evaluation. Users can seamlessly work with both open models (Llama, QWEN, Phi and others) and commercial APIs (OpenAI, Anthropic, Vertex AI), with both text and multimodal models. Oumi will offer prebuilt ready-to-use workflows and recipes for post training and other common operations.

Oumi can run anywhere so that developers can train and evaluate models seamlessly across environments, from local machines to remote clusters and clouds (AWS, Azure, GCP, Lambda), with native support for Jupyter notebooks and VS Code debugging. Additionally, the enterprise-grade platform is built for scale, offering first-class support for distributed training with PyTorch DDP and FSDP and ability to efficiently handle models up to 405B parameters.

To better enable the community to build upon each other's work independently, or to collaborate in research efforts, the recordability and reproducibility of experiments is a key design principle of the Oumi platform. To further promote the collective improvement of foundation models, Oumi plans to coordinate joint research efforts with fully open participation across the world.

Oumi is developed in partnership with researchers at the following universities:

  • University of Illinois Urbana Champaign
  • Carnegie Mellon University
  • Princeton University
  • Stanford University
  • Georgia Institute of Technology
  • California Institute of Technology
  • University of California, Berkeley
  • University of Washington
  • New York University
  • Massachusetts Institute of Technology
  • University of Waterloo
  • University of Cambridge
  • University of Oxford

Developers and researchers interested in contributing to building towards a better AI future should join the Oumi community by visiting the Oumi website or GitHub page.

Oumi AI
Built to be truly open for everyone
Truly open and collectively developed AI is the future.
Build that future.