How a free software strategy catapulted DeepSeek to AI stardom
Summary
- The Chinese AI upstart made the shrewd bet that American developers will latch on to its technology because it is open-source.
DeepSeek, the Chinese artificial intelligence company that set Silicon Valley ablaze with its new AI model, did it in part by relying on a decades-old strategy: giving its technology away free.
The approach, known as open-source, relies on companies freely distributing their technology for the public to use, share and modify. It has long been used as a tactic by upstarts to upset the established dominance of larger, closed-source competitors that don’t freely share their tech.
In this case, it’s DeepSeek going up against OpenAI, maker of the ChatGPT chatbot. By some analyst estimates, OpenAI accounts for 80% of the global generative AI market. DeepSeek said its model was particularly good at problem solving, performing on par with OpenAI’s o1 reasoning model—but at a fraction of the cost per use.
It’s not just DeepSeek using the open-source strategy to challenge OpenAI. AI startups including Mistral AI, Hugging Face, Runway AI and Together AI have similarly released their technology as open-source. Tech giants have done it too: Meta Platforms open-sourced its Llama model in 2023, and Google released its Gemma open-source models last year.
Still, it is a matter of debate whether DeepSeek’s R1 model is actually “open-source" by standard open-software definitions, said Arun Chandrasekaran, an analyst focused on AI at market research and IT consulting firm Gartner.
While DeepSeek has released the “weights," or numerical parameters, behind its R1 model for the public to freely use, download and modify, it stopped short of releasing training data behind it. U.S. companies like Meta have taken the same approach, while calling their models “open-source."
Yet being free means open-source technologies can rapidly capture the attention of legions of developers eager to experiment with new tech. And in doing so, they create a so-called ecosystem of engineers who build their own versions of the technology—helping it proliferate well beyond its original builders’ designs.
Hugging Face, a startup that hosts a popular open-source model-sharing platform, said that 617 models have already been created by its community based on DeepSeek’s R1 model, with over 3 million downloads.
“Openness raises all tides when it comes to AI development," said Jason Corso, a University of Michigan professor of robotics. “They’ve enabled the broader AI community to study and verify their methodology and weave it into their work."
If DeepSeek’s new R1 model were proprietary, or pay-to-use, independent developers would be less likely to expend their efforts to build and share improvements they’ve made to the technology. Winning over developers is still key to gaining a foothold in AI, especially when AI companies like OpenAI and Anthropic continue battling for engineers to build applications on their technologies.
Startups that build on open-source technologies are also more likely to experiment with powerful, cheaper models like DeepSeek’s—helping embed the Chinese company’s AI model into new products and services that touch consumers and businesses.
American companies have also expressed excitement over DeepSeek’s capabilities, with corporate technology executives saying they hope that it brings down the cost of AI for everyone.
“There will be another DeepSeek by the end of this month, and many more this year in the U.S. and worldwide," said Jim Zemlin, executive director of the Linux Foundation, a nonprofit that supports open-source technologies.
One reason companies lock down their technology is to prevent competitors from accessing it. But even when technologies aren’t shared freely, it can be tricky to stop their dissemination. OpenAI on Wednesday said it was investigating whether DeepSeek trained its new chatbot by repeatedly querying OpenAI’s models.
There’s no doubt that open-source can be a risky approach. There’s no guarantee that free technology will turn a profit, leading some companies to sell business-grade services and applications on top of their open models. Some also sell access to their most powerful models, while giving away less capable models free.
Open-source technology can also be a red-flag for enterprises, who are wary of bringing unvetted technology into their premises. Experts say that open technologies make it easier for attackers to identify and exploit potential cybersecurity issues, but open-source also offers the benefit of greater customization and control.
“Open-weight models lack built-in security certifications, placing the burden of compliance on the deploying organization," said Ryan Cox, global head of AI at Synechron, a technology and consulting firm focused on financial services.
But for DeepSeek, a goal may simply be to maximize its popularity and demonstrate to America’s AI leaders that it has developed a powerful, cheaper technology.
“They realized they were onto an extremely powerful model, and they said, ‘This is our moment, let’s make this really viral,’" Gartner’s Chandrasekaran said. “One of the ways to make it really viral is by open-sourcing it."
Write to Belle Lin at belle.lin@wsj.com
