Be a part of our every day and weekly newsletters for the most recent updates and unique content material on industry-leading AI protection. Study Extra
If it wasn’t clear earlier than, it’s undoubtedly very clear now: Open supply actually does matter for AI. The success of DeepSeek-R1 has substantively confirmed there’s a want and demand for open-source AI.
However what precisely is open-source AI? For Meta and its Llama fashions, it means free entry to make use of the mannequin, with some situations. DeepSeek is out there below a permissive open-source license with the mannequin code open and out there for anybody to make use of. What neither method allows, nevertheless, is full unconditional entry to all of the mannequin code, together with weights in addition to coaching information. With out all that info, builders can nonetheless work with the open mannequin however they don’t have all the mandatory instruments and insights to grasp the way it actually works and extra importantly tips on how to construct a wholly new mannequin. That’s a problem {that a} new startup led by former Google and Apple AI veterans goals to unravel.
Launching in the present day, Oumi is backed by an alliance of 13 main analysis universities together with Princeton, Stanford, MIT, UC Berkeley, College of Oxford, College of Cambridge, College of Waterloo and Carnegie Mellon. Oumi’s founders raised $10 million, a modest seed spherical they are saying meets their wants. Whereas main gamers like OpenAI ponder $500 billion investments in huge information facilities by means of tasks like Stargate, Oumi is taking a radically totally different method. The platform offers researchers and builders with an entire toolkit for constructing, evaluating and deploying basis fashions.
“Even the most important corporations can’t do that on their very own,” Oussama Elachqar, cofounder of Oumi and beforehand a machine studying engineer at Apple, informed VentureBeat. “We have been successfully working in silos inside Apple, and there are various different silos taking place throughout the {industry}. There needs to be a greater method to develop these fashions collaboratively.”
What open-source fashions like DeepSeek and Llama are lacking
Oumi CEO and former Google Cloud AI senior engineering supervisor Manos Koukoumidis informed VentureBeat that researchers persistently inform him AI experimentation has turn out to be extraordinarily complicated.
Whereas in the present day’s open fashions are a step ahead, it’s not sufficient. Koukoumidis defined that with present “open” AI fashions like DeepSeek-R1 and Llama, a company can use the mannequin and deploy it on their very own. What’s lacking is that anybody else who desires to construct on the mannequin doesn’t know precisely the way it was constructed.
The Oumi founders consider this lack of transparency is a significant hindrance to collaborative AI analysis and improvement. Even a undertaking like Llama requires a big quantity of effort from researchers to determine tips on how to reproduce and construct upon the work.
How Oumi works to open AI for enterprise customers, researchers and everybody else
The Oumi platform works by offering an all-in-one setting that streamlines the complicated workflows concerned in constructing AI fashions.
Koukoumidis defined that to construct a basis mannequin, there are sometimes 10 or extra steps that have to be finished, typically in parallel. Oumi integrates all obligatory instruments and workflows right into a unified setting, eliminating the necessity for researchers to piece collectively and configure varied open-source parts.
Key technical options embrace:
- Help for fashions starting from 10M to 405B parameters
- Implementation of superior coaching strategies together with SFT, LoRA, QLoRA and DPO
- Compatibility with each textual content and multimodal fashions
- Constructed-in instruments for coaching information synthesis and curation utilizing LLM judges
- Deployment choices by means of trendy inference engines like vLLM and SGLang
- Complete mannequin analysis throughout normal {industry} benchmarks
“We don’t should cope with the open-source improvement hell of determining what you’ll be able to mix and what works effectively,” Koukoumidis defined.
The platform permits customers to start out small, utilizing their very own laptops for preliminary experiments and mannequin coaching. As customers progress, they’ll then scale as much as bigger compute sources, resembling college clusters or cloud suppliers, all throughout the identical Oumi setting.
You don’t want huge coaching infrastructure to construct an open mannequin
One of many large surprises with DeepSeek-R1 is the truth that it was apparently constructed with a fraction of the sources that Meta or OpenAI use to construct their fashions.
As OpenAI and others make investments billions in centralized infrastructure, Oumi is betting on a distributed method that would dramatically scale back prices.
“The concept you want lots of of billions [of dollars] for AI infrastructure is basically flawed,” Koukoumidis stated. “With distributed computing throughout universities and analysis establishments, we are able to obtain related or higher outcomes at a fraction of the fee.”
The preliminary focus for Oumi is to construct out the open-source ecosystem of customers and improvement. However that’s not all the corporate has deliberate. Oumi plans to develop enterprise choices to assist companies deploy these fashions in manufacturing environments.
