How small open-source AI startups are taking on the big developers

When the Chinese AI startup DeepSeek wiped billions off the market value of the largest US tech companies earlier this year with the release of its cheap and powerful chatbot model, it confirmed for Pierre-Carl Langlais what he already knew: the future of AI belongs to small models.

Langlais is an AI researcher and the founder of French startup Pleias, which is training small, open-source AI models designed to specialise in document and data processing. His company is part of a wave of startups that consider themselves a challenge to the dominance of the large developers like OpenAI. He says DeepSeek’s focus on open-source development – where the model’s workings are freely shared for anyone to use – was a shrewd bet.

“It’s actually very powerful because it means at some point you are the one setting standards, you are the one which is going to be the reference [for other AI developers],” he says.

Langlais believes one of the main ways AI will generate value is through training on a business customer’s proprietary data to create a specialised model. The AI researcher and investor Andrew Ng recently made this point on stage at the GCVI Summit, saying that most large language models like the one underpinning ChatGPT are trained on publicly available datasets, but that a lot of untapped valuable data, like company emails and records, is not published on the internet.

This shift to a more bespoke application of AI is already happening. OpenAI’s GPT-4o model allows companies to upload their data to fine-tune the model for tasks specific to their needs. But Langlais thinks this is where small models made by “very agile startups” can compete with the leading large language model developers like OpenAI, Microsoft, Google and Anthropic.

China’s DeepSeek showed that it was possible to build a comparable model to OpenAI’s on what is believed to be a much lower training cost. Langlais thinks that the cost of training is only going to fall further, and training methods will become more widely understood, especially with developers like DeepSeek releasing their models on open licences. Corporates looking for bespoke AI solutions will then have more choice: they could hand over all their proprietary data to a large AI company’s API, or they could work with a smaller startup developer to build a model that they might have more control over.

Langlais thinks this trend will only deepen now that AI’s main use in business is expected to evolve from chatbots to AI agents, which are capable of autonomously performing tasks. Smaller, specialised models could be an efficient way of powering them.

“We’re going to see a lot of disruptive breakthroughs,” he says. “I think in the coming months we are going to be very surprised.”

But a startup wanting to take on the largest AI labs has quite a challenge ahead of them. OpenAI, for example, recently announced a funding round that could mean as much as $40bn will be invested, if all conditions are met. Who would want them as a competitor? Or, for that matter, the likes of Google, Microsoft and xAI?

Programmers of the world unite

Alex Ferguson, head of growth for the AI startup Prime Intellect, says that his company was built because the founders believe advanced AI technology should not be concentrated among the powerful AI labs with their proprietary models.

“What if one party owned all this intelligence, and this thing that could potentially change all of our lives drastically?” he says.

“What if one party owned that and could outsmart anything or anyone if they wanted to try to get control of it? That leads to a lot of interesting and scary questions.”

He describes Prime Intellect as a decentralised AI company. It pools compute from different sources, such as individual programmers, potential users or enthusiasts, to build an open-source model. Anybody, in theory, can contribute, and anybody can use the results.

“You could have, say, mutual parties that are interested in the same ideal outcome, and they could come together and do a distributed, decentralised training on that ideal outcome as a community and share the ownership of that,” he says.

“That’s never been done before.”

He gives a hypothetical example of SaaS companies banding together to build a shared model that can create agents for their CRM systems.

But for now, with over $15m in funding raised from investors that include VC firm Founders Fund and the open-source research and development company Protocol Labs, Ferguson says the priority is developing the overall technology, rather than immediately focusing on one business area.

Prime Intellect recently released its second model, Intellect-2, trained using crowdsourced compute. It also compiles and shares synthetic datasets. The company’s website says it aims to build a decentralised model that rivals the frontier of AI development for performance. According to Ferguson, the ultimate “north star” is developing open-source artificial general intelligence (AGI), a contested term that typically refers to a state of AI technology where it can outperform humans on most if not all cognitive tasks.  

But in the more immediate future, the technology will probably be most impactful when used as Langlais expects. If distributed sources of compute make high-quality AI models more accessible, then it could lead to a proliferation of small, specific uses.

“I think there’s so many different use cases where you can fine tune things, make them more specific to an ideal outcome,” says Ferguson. “And I feel like we’ve seen a lot of value creation by doing this within the small langue model world, within the agent world.”

Open-source AI vision

No AI developer is ever likely to think up all the possible uses of their technology. Jay Allen, CEO of Moondream, a US vision AI startup, says that having an open-source product has meant users have been able to come up with uses for Moondream’s technology that the company never would have thought of.

“As a result, they come back to us and say, ‘Hey, I own a cattle ranch in Texas, and once in a while I lose a cow, and I’m using a drone now to go look around and see if it’s a cow, and Moondream is really good at that,’” says Allen.

“We never would have built a cow-finding drone company. That never would have happened.”

Moondream makes vision AI models called vision language models (VLM), which resemble LLMs like OpenAI’s GPT range in the sense that they can receive, process and produce text, but which are also trained with images and video for visual analysis and reasoning. It received pre-seed funding from investors that included Microsoft’s M12 GitHub Fund, which invests in open-source software.

The technology does not focus on image generation, like some VLMs. Instead, it is to help computers see. It advertises features including gaze detection, object recognition and document reading.

Allen says that while Moondream may be at a disadvantage to the large AI labs in terms of brand recognition, he thinks the company can eventually come to compete on price, because it is a smaller model and so has lower inference costs.

“Every video frame is going to cost me exponentially more [in a large model] than something like Moondream, a small 2B [two billion parameter] model,” he says.

But this still means the model has to perform well. Allen compares Moondream to DeepSeek, as an AI company that is “achieving really good results despite having a fraction of the spend.”

He attributes this in part to the company’s commitment to using only the highest quality data, with each dataset checked thoroughly for errors, which he says are common from third-party sources. They are also careful about what they teach the model, to ensure there is no waste of resources.

“Our model does not know what Kim Kardashian looks like. Our model does not know world history,” he says.

Like DeepSeek, Pleias and Prime Intellect, Moondream lacks the vast computational resources of the large AI developers. The crucial skill has been to do more with less. Allen says that the lack of compute Moondream started with has, counter-intuitively, helped them develop a stronger model, because training iterations are shorter and cheaper, allowing them to experiment more.

And Allen believes the open-source community of AI developers will drive widespread development of models.

“If one model becomes really good at something, it becomes relatively easy for another model to catch up,” he says. “And so being open-source and making it easier and lowering those barriers means everybody’s moving faster.”

Whether these smaller companies can mount a threat to the larger AI developers depends on the quality, price and use cases of the models they produce. But taking DeepSeek as an inspiration, they all believe it is possible.