If it wasn’t clear before, it’s definitely very clear now: Open source really does matter for AI. The success of DeepSeek-R1 has substantively proven there is a need and demand for open-source AI.
But what exactly is open-source AI? For Meta and its Llama models, it means free access to use the model, with some conditions. DeepSeek is available under a permissive open-source license with the model code open and available for anyone to use. What neither approach enables, however, is full unconditional access to all the model code, including weights as well as training data. Without all that information, developers can still work with the open model but they don’t have all the necessary tools and insights to understand how it really works and more importantly how to build an entirely new model. That’s a challenge that a new startup led by former Google and Apple AI veterans aims to solve.
Launching today, Oumi is backed by an alliance of 13 leading research universities including Princeton, Stanford, MIT, UC Berkeley, University of Oxford, University of Cambridge, University of Waterloo and Carnegie Mellon. Oumi’s founders raised $10 million, a modest seed round they say meets their needs. While major players like OpenAI contemplate $500 billion investments in massive data centers through projects like Stargate, Oumi is taking a radically different approach. The platform provides researchers and developers with a complete toolkit for building, evaluating and deploying foundation models.
“Even the biggest companies can’t do this on their own,” Oussama Elachqar, cofounder of Oumi and previously a machine learning engineer at Apple, told VentureBeat. “We were effectively working in silos within Apple, and there are many other silos happening across the industry. There has to be a better way to develop these models collaboratively.”
What open-source models like DeepSeek and Llama are missing
Oumi CEO and former Google Cloud AI senior engineering manager Manos Koukoumidis told VentureBeat that researchers consistently tell him AI experimentation has become extremely complex.
While today’s open models are a step forward, it’s not enough. Koukoumidis explained that with current “open” AI models like DeepSeek-R1 and Llama, an organization can use the model and deploy it on their own. What’s missing is that anyone else who wants to build on the model doesn’t know exactly how it was built.
The Oumi founders believe this lack of transparency is a major hindrance to collaborative AI research and development. Even a project like Llama requires a significant amount of effort from researchers to figure out how to reproduce and build upon the work.
How Oumi works to open AI for enterprise users, researchers and everyone else
The Oumi platform works by providing an all-in-one environment that streamlines the complex workflows involved in building AI models.
Koukoumidis explained that to build a foundation model, there are typically 10 or more steps that need to be done, often in parallel. Oumi integrates all necessary tools and workflows into a unified environment, eliminating the need for researchers to piece together and configure various open-source components.
Key technical features include:
- Support for models ranging from 10M to 405B parameters
- Implementation of advanced training techniques including SFT, LoRA, QLoRA and DPO
- Compatibility with both text and multimodal models
- Built-in tools for training data synthesis and curation using LLM judges
- Deployment options through modern inference engines like vLLM and SGLang
- Comprehensive model evaluation across standard industry benchmarks
“We don’t have to deal with the open-source development hell of figuring out what you can combine and what works well,” Koukoumidis explained.
The platform allows users to start small, using their own laptops for initial experiments and model training. As users progress, they can then scale up to larger compute resources, such as university clusters or cloud providers, all within the same Oumi environment.
You don’t need massive training infrastructure to build an open model
One of the big surprises with DeepSeek-R1 is the fact that it was apparently built with a fraction of the resources that Meta or OpenAI use to build their models.
As OpenAI and others invest billions in centralized infrastructure, Oumi is betting on a distributed approach that could dramatically reduce costs.
“The idea that you need hundreds of billions [of dollars] for AI infrastructure is fundamentally flawed,” Koukoumidis said. “With distributed computing across universities and research institutions, we can achieve similar or better results at a fraction of the cost.”
The initial focus for Oumi is to build out the open-source ecosystem of users and development. But that’s not all the company has planned. Oumi plans to develop enterprise offerings to help businesses deploy these models in production environments.
Published by: Sean Michael Kerner
