Two months on from its release, DeepSeek’s R1, which wowed experts and caused American tech stocks to crash in January, is still unbeaten. The Chinese firm’s artificial-intelligence (AI) model remains the best open-source offering released by any lab, anywhere in the world. Back home it has spread like wildfire. It is catching on in the rest of the world as well.
The value offered by DeepSeek’s open-source models is hard to ignore. They can be used by coders without asking permission or paying a fee, and can be downloaded and run on a company’s own hardware if it wishes to do so for performance or privacy reasons. That puts DeepSeek in a different category to more technically impressive but closed labs like OpenAI.
Some companies in the West have already begun to adopt DeepSeek’s models, which are now supported by cloud providers including Amazon and Microsoft. Gloo, a messaging platform for churches founded by Pat Gelsinger, who until December was the chief executive of Intel, has used the model as the basis for its chatbot. Latenode, an automation platform, began offering R1 shortly after the model was launched, noting that its support for non-English languages was appealing to content marketers looking to translate material. Meta, DeepSeek’s main open-source rival, has assigned researchers to pull apart R1 and apply the lessons to its own family of Llama models.
Perhaps surprisingly, European banks have emerged as a hotbed of experimentation. Strict confidentiality rules limit how much the financial sector can rely on cloud-based AI services. That makes open-source models hosted internally an attractive alternative. Natwest and HSBC, two British lenders, are both experimenting with building their services on top of R1, as is Spain’s BBVA, according to the Information, a news site.
Many Western businesses, though, remain cautious. Security is often the concern, and one that may preoccupy Americans more than Europeans. One American boss says he would only run DeepSeek’s models on an “air-gapped” computer with no connection to his firm’s systems—even though it would have taken a breakthrough in computer science for DeepSeek to have smuggled malicious capabilities into the model itself.
Moreover, the White House is considering banning DeepSeek’s chatbot app on government devices, on national-security grounds, and could decide to go further, including possibly preventing American cloud providers from offering DeepSeek’s models at all. That would have a chilling effect on adoption.
Europeans, by comparison, may find themselves with little choice. None of the old continent’s homegrown ai champions are of the calibre of DeepSeek. Donald Trump’s trade war and his overtures towards Vladimir Putin mean that some European companies and policymakers may want to lower their reliance on American tech.
In late February DeepSeek made it even easier for others to harness its technology by making the code it used to create its models available free of charge. Previously DeepSeek had only shared details of the process it used to train its systems, which was far more efficient than the approaches taken by Western firms. But now companies have access to the exact code used by the Chinese pioneer to develop its models. Thousands have downloaded it in the past week alone.
Using DeepSeek’s methodology may be a good option for companies looking to benefit from its breakthroughs while avoiding the Chinese censorship that is embedded in its models—a final hurdle to adoption. Ask DeepSeek’s chatbot about Tiananmen Square, for example, and it will want to “talk about something else”. Others have taken a more direct approach to tackling that problem. Perplexity, an American AI search engine, has taken R1 and retrained it to ensure it produces “unbiased, accurate, and factual information”. It has dubbed its creation, with patriotic fervour, “R1 1776”. ■
To stay on top of the biggest stories in business and technology, sign up to the Bottom Line, our weekly subscriber-only newsletter.