Open Source AI: The DeepSeek Takeaway for Europe

On January 20th, DeepSeek sent shockwaves through the AI community by releasing models matching or exceeding OpenAI's frontier models on reasoning tasks, while demonstrating how to distill advanced capabilities into smaller, efficient models. The most impressive part? DeepSeek's reported model training costs were just $5.6 million compared to OpenAI's reported $78 million for ChatGPT-4o.

While many focus on the US-China AI race or interpret DeepSeek as a “Sputnik moment”, they're missing the real story. This isn't about China versus the USA (or Europe!) – it's about the fundamental power of open innovation to reshape markets and facilitate collective progress and community-driven collaboration at the frontiers of science and technology. As Nathan Lambert from the Allen Institute for AI succinctly put it: DeepSeek is an example of “the [open source] system working as intended.” And crucially for Europe, DeepSeek’s models highlight that the AI race is far from over and Europe is far from being shut out of this race. We can compete, dare I say lead in this race, if we double down on open source and become a leader in open source AI.

DeepSeek's achievement demonstrates that AI competitiveness isn't just about enormous capital expenditure on chips and compute, but also ingenuity. As the DeepSeel developers explain in their technical report, they were able to produce models that achieved performance on par with frontier proprietary models through the novel application of reinforcement learning to open models, letting the model explore chain-of-thought reasoning through trial-and-error before combining this with supervised training. The result was models that match proprietary leaders but at a fraction of the cost, at least according to reported costs. Going further, they showed that the advanced capabilities of large models like DeepSeek-R1 (671B parameters) can be distilled into much smaller but nonetheless extremely powerful models and released distilled models of different sizes (1.5B to 70B) based on Alibaba’s Qwen 2.5 and Meta’s Llama 3 open models.

Crucially, what DeepSeek accomplished wasn't magic. There were no massive algorithmic breakthroughs – just clever engineering and a return to first principles. They stacked existing techniques like reinforcement learning and fine-tuning in novel ways and achieved outsized results. The implications of this innovation cannot be overstated: they showed that breakthrough innovations in AI development don't always require massive capital investments or groundbreaking new theories – sometimes they just need a fresh perspective and clever engineering, a metaphorical reshuffling of the cards. 

This brings competitiveness back to the realm of clever engineering rather than just massive investments, democratizing access to and the development of state-of-the-art AI models.

Europe is uniquely positioned to leverage this opportunity. With our robust ecosystem of AI talent and history of open innovation (both the Internet and Linux were created here), European developers are already leaders in the open source AI community. Companies like Mistral AI, Aleph Alpha, and Black Forest Labs are gaining ground, while initiatives like EuroLLM are producing competitive multilingual models for all 24 official European languages.

Europe's approach to AI can be distinctly different, leveraging our strengths in transparency, privacy, and responsible development. For example, while DeepSeek's models impress, the community have noted DeepSeek’s models avoid discussing topics censored in China. This is not the kind of AI we want in our society and I commend researchers at Perplixity AI for uncensoring DeepSeek’s models (and showing that this is possible, too!).

Europe can lead with models that reflect our values while remaining competitive. The advantage is twofold: European companies can focus on solving real-world problems in our society and our sectors rather than competing to build foundation models that chase after artificial general intelligence, and by building applications on top of models that can align with EU values and regulations from the start. In fact, Nick Clegg from Meta made this point at the World Economic Forum last month: building AI applications on top of models is an area where Europe can develop a competitive edge. To do so competitively, we need to drive down the cost of model inference and the growing momentum of small but powerful models gives us reason to be optimistic.

Recently, we’ve heard of a number of grand investments in AI infrastructure. For example, at the AI Action Summit, President Macron announced a €109 billion plan for AI compute infrastructure, matching the recently announced US Stargate program per capita.  It is great to see this ambition in Europe. As per the Draghi report, Europe must make unprecedented economic investment in digital innovation if it wants to be competitive in the global economy. 

But I want to highlight that the key to success isn't just about money – it's about how we invest it strategically across both the public and private sectors. And open source cannot be just a part of Europe’s digital strategy; it’s the only way for Europe to achieve its goals of competitiveness and promotion of its democratic values at a global scale, especially in the era of AI. Specifically, our investment strategy must support our digital commons that is key to AI research and development — the open source software, open data, and open models that are developed, maintained, and governed by global online communities.

More specifically, we should be investing in the sustainability of the open source developer communities that have given us invaluable libraries like scikit-learn and PyTorch without which AI would simply not be possible.  We should also invest in open source AI safety tooling that facilitate responsible and ethical development. I point to examples like the UK AI Security Institute's Inspect framework for LLM evaluations, ETH Zurich's Compl-AI and the newly formed ROOST initiative. And we must invest in the creation of a strong commercial ecosystem with favorable conditions for open source businesses to thrive, grow, and exit. 

It goes without saying that data is key to AI and as such we should also be investing more in the communities that develop, maintain, and govern our open data commons. For example, in Europe we could invest more in the creation and curation of open data that capture Europe's rich linguistic and cultural diversity, from the EU’s 24 official languages to regional dialects, or ones that facilitate innovation in industries and sectors where Europe has a competitive advantage like automotive or renewable energies. 

This is all to say: DeepSeek's success hasn't closed doors for Europe – it's opened them. By embracing open source, Europe can lead in AI development that reflects our values while driving innovation. This isn't about catching up to the US or China – it's about leading the world toward a more open, transparent, and democratic future.