Making Good on the Promise of Open Source AI (original) (raw)
The AI industry has seemingly been trying to define what “open” and “open source” mean in the generative AI era that kicked off more than two years ago when OpenAI released its ChatGPT chatbot and has been evolving rapidly since.
The AI space is dominated by a relatively few large and well-funded companies and their AI models and tools, such as OpenAI’s GPT family of models, Microsoft’s Copilot, and Google’s Gemini. Others, with Meta being the most high-profile of the group with its Llama models, boast of being open, though there are detractors who question how open they really are.
However, beyond the debate about definitions, the lack of real and widely used open source AI platforms is hindering innovation in the industry and creating the illusion of a troubling talent gap, according to Manos Koukoumidis, who has worked on AI technology at such heavyweights as Google, Microsoft, and Meta.
“Quite often people, they release open weight models, just the model itself and the weight,” Koukoumidis told The New Stack. “[People] don’t know exactly how it was developed, what code, what data was used and they call it open source, which is not a very accurate description. The reality is that, even with current efforts, even with the few ones that are open source — because Meta’s ‘open weight’ — even for the few that are open source, it’s very hard for people to experiment, to continue to innovate and collaborate with each other on this work. This is the thing that is impeding progress in the open source world.”
It’s what led Koukoumidis and Oussama Elachqar — who brings machine learning experience with Microsoft, Apple, and Twitter (now X) to the table — to launch Oumi, a startup coming out of stealth today that offers what both call a true open source platform. It also has $10 million in seed funding, led by capital venture firms Venrock and Obvious Ventures.
Linux for AI
Koukoumidis, who also is Oumi’s CEO, said the platform — developed in collaboration with researchers from 11 institutions, including Massachusetts Institute for Technology (MIT), University of California, Berkely, and Carnegie Mellon University — essentially acts as the Linux for AI models and tools, enabling wide collaboration and contributions researchers, developers, and AI experts, who will be able to build off each other’s work.
It will not only accelerate innovation but also will give AI students at such institutions the ability to work with the platform to gain the skills that are so badly needed in the industry. It helps open up the critical advanced technology to more than the current handful of power brokers.
“We started with the ability for others to experiment easily and collaborate as the key design principle,” he said. “If it’s not easy for others to build on each other’s work and continue building on top of the work of each other, then open source will never achieve the growth and the velocity that it needs.”
Making AI Really Open
Making AI technology more open has been a thorny issue in the industry. As mentioned, companies like Meta are creating models that are more open than some of their competitors. The Open Source Initiative, after years of planning, in October 2024 introduced its initial definition of open source AI, which addresses four different kinds of data and requires those building and sponsoring AI technology to share what data they can as well as the model’s parameters and the source code used to train and run the system.
With the rise of agentic AI, Cisco and other vendors are mapping out communication networks — what they call the “Internet of Agents” — that will allow AI agents from different vendors and industries to autonomously connect and collaborate to solve complex problems, and are stressing the need for it to be open source.
Enter DeepSeek
More recently, China AI startup DeepSeek this month released its DeepSeek-R1 reasoning model, which is roiling the AI industry because it reportedly equals the performance of OpenAI’s o1 model — which can reason before giving an answer — and was trained in two months for just $5.6 million. It’s also open source, so AI developers and researchers can collaborate and build upon it.
After the R1 release, stocks of Nvidia, Microsoft, and other AI giants spiraled and DeepSeek knocked OpenAI from its perch as the most downloaded free app on Apple’s App Store.
“The release of DeepSeek undeniably showcases the immense potential of open- ource AI,” said Andrew Bolster, senior R&D manager at secure development firm Black Duck. “By making such a powerful model available under an MIT license, it not only democratizes access to cutting-edge technology but also fosters innovation and collaboration across the global AI community.”
However, DeepSeek’s use of OpenAI’s Chain of Thought data for initial training puts a spotlight on the need for transparency and shared resources, Bolster said, adding that “it’s crucial that the underlying training and evaluation data are open, as well as the initial architecture and the resultant model weights.”
DeepSeek Helps Make the Open Source Argument
Koukoumidis said DeepSeek’s success validates Oumi’s strategy. Being built atop other open source efforts like Meta’s Llama and PyTorch enabled the Chinese company to innovate and create models that seem to be closing the gap with proprietary models from OpenAI.
DeepSeek was built on top of other open source efforts like PyTorch and Llama from Meta. Those open source efforts enabled the DeepSeek team to continue innovating and develop models that appear to be closing the gap with the proprietary models from OpenAI. On a platform like Oumi’s, the community can co-build the next DeepSeek, he said.
Koukoumidis said DeepSeek also should serve as a warning to the United States.
“Despite its impressive performance in areas like math and coding, DeepSeek’s biases and censorship indicate what is at stake for the future of AI development,” he said. “For the U.S. to continue to lead in AI, we need open source and open collaboration to develop trustworthy and explainable models. U.S. researchers are at a disadvantage if the U.S. is less open than China.”
Communicate and Collaborate
Koukoumidis and Elachqar call Oumi an AI laboratory that is rolling out the first unconditionally open source AI platform. It provides foundation models with open code, open data, and open weights and allows researchers and developers with the tools to collaborate and contribute. It is a unified platform that can support all common foundation model workflows.
Developers can use techniques like SFT, LoRA, QLoRA, and DPO to train models of various sizes, from 10 million to 405 billion parameters, get support for PyTorch and other AI tools, work with text and multimodal models like Llama, Qwen (the LLM family built by Alibaba Cloud, and Microsoft’s Phi small language models (SLMs).
They can use multiple inference engines like vLLM and SGLang, evaluate models via standard benchmarks, and run their models in any environment, from their own laptops to cloud infrastructures from Amazon Web Services and Microsoft’s Azure to Google Cloud Platform and Lambda. In addition, developers can integrate their models with open model or commercial APIs from OpenAI, Anthropic, Google’s Vertex AI, and others.
There also is native support for Jupyter notebooks and Microsoft Visual Studio code debugging. Oumi also includes prebuilt workflows and recipes for various operations, including post-training.
“When there’s a new project, when they have a research idea, they want to execute on it,” Elachqar told The New Stack. “But there’s so much that they have to figure out just to test that idea, the hypothesis. What we provide them is for the most common workflows, what they need to fine-tune a model or generate data or to inference. We provide them with really solid starting points that they can use and tweak to their use case.’
Betting on the Right Horse
It’s crucial that the industry moves toward true open source and away from one-off AI models built by companies behind the wall, the Oumi founders said. Innovation will happen faster and there will be more available talent to draw on, and they believe players in the ecosystem — from the cloud providers to accelerator vendors on down to small companies and research institutions — all want open source AI to succeed.
The most resourced companies can’t solve challenges alone, Elachqar said, noting the years of experience both he and Koukoumidis have with such vendors. Open collaboration is essential; otherwise, AI developers are working in silos and creating the same tools that everyone else is rather than building off what has already been done.
It also will protect the industry itself. Koukoumidis expects many companies making closed models to collapse in the next few years. He pointed to Inflection AI, a startup that until late last year was rolling out models that executives said could challenge the leaders. It’s now gotten out of the model-building game, shifting to creating tools that enterprises can use with AI models built by others.
“You have two horses in the race,” the CEO said. “There’s the closed-source horse that says, ‘I need to do everything by my own. I need to do all the effort by my own to develop and I have to get the full cost of doing this.’ And then you have an open source horse that so many people are providing resources to. They’re helping it move faster, and it’s aggregating all the contributions — both human and monetary contributions — from all the different entities. And the question is, which one do you think is going to be faster and more economically sustainable in the end? We are betting on the second horse.”
Group Created with Sketch.