Tech companies and academics have long wrestled with the risks and rewards of building open source software. But the frenzy around generative artificial intelligence has lent new significance to the debate.

The rise of the Chinese AI startup DeepSeek has only added attention to the issue. DeepSeek recently released an open source model that it said rivaled software from the top American AI developers — and it claimed to have done so for a fraction of the development cost, using less powerful hardware.

Here’s what you need to know.

What is open source?

In general, open source refers to software with source code that is made freely available for anyone to use or modify. According to the Open Source Initiative (OSI), a California-based nonprofit that advocates for the benefits of this way of working, to qualify as open source, software must comply with specific terms for distribution and access.

For an AI model specifically to be considered truly open, the OSI says its developers must provide detailed information about its training data and give people the ability to study, use and modify the system for any purpose.

Closed source, by contrast, refers to software and models that are controlled by the developer, can’t be modified, and offer less transparency about their technical underpinnings.

Do any top AI developers offer open source software?

Many tech companies brand their AI software as open source; not everyone agrees they all live up to that definition.

Meta Platforms Inc., French startup Mistral and now DeepSeek have all released AI models that they call open source. (OpenAI, despite having “open” in its name, does not open source most of its models.)

But often such models are actually what are known as open weight models. That means that in addition to offering up the model, and perhaps some of its source code, the companies disclose the weights — that is, the many numerical values the model picked up and was tweaked on during its training process, which allows developers to better customize it — but not details about the data actually used to train it.

Meta, for example, offers weights and some of the source code for its Llama series of AI models, but does not provide detailed information about its training data. Meta has also previously been called out by the OSI for licensing terms that include certain restrictions on commercial uses.

Similarly, in January, DeepSeek said it released its latest system, R1, as an open source model, but it did not offer code or training data. That led to questions about what, exactly, the company may have used to build its model.

What are the benefits of open source?

Those who promote open source software typically tout it as being more affordable for users because it doesn’t have the same licensing fees. Cheaper prices pave the way for broader AI adoption, and reduced development costs promote innovation. Supporters also note that the approach boosts accountability for developers creating powerful AI systems by giving others the ability to better understand how the models work.

Closed systems pose the risk of producing an AI market dominated by a handful of powerful companies. Aaron Levie, chief executive officer of cloud storage company Box Inc., recently said: “In the world of very expensive and proprietary AI, the providers of AI could and likely should choose to keep all the economics for themselves — basically crowding out opportunity for developers and the ecosystem.”

For companies such as Meta, there’s an added benefit to open source: popularity. By allowing other developers to freely access and build on top of its open source software, Meta has been able to expand its influence throughout the AI ecosystem.

What are the risks of open systems?

Critics of open source software argue that it’s less secure. In the case of AI, some in the U.S. fear that using such models from geopolitical rivals such as China pose a risk to national security, threatening citizens’ safety — such as by collecting massive amounts of user data that could be used for surveillance purposes. And there are concerns that American companies offering their AI models with varying degrees of openness may potentially give rival countries an opportunity to use them to one-up U.S. technological dominance.

Why did DeepSeek go the “open” route?

By embracing a more open approach (with some caveats), DeepSeek may have eased some concerns among global users about China’s tight control of the technology. The startup also likely broadened the reach of its chatbot in Western markets by making it easier for other developers to adapt the underlying technology to meet their needs.

In other words, DeepSeek effectively followed the same playbook Meta has used to capture more of the AI ecosystem — a fact that seems not to have gone unnoticed by Meta CEO Mark Zuckerberg.

“This is a huge geopolitical competition, and China’s running at it super hard,” Zuckerberg said in an interview on The Joe Rogan Experience. “If there should be an open source model that everyone uses, we should want it to be an American model.”

So how does DeepSeek’s model work?

Like some of the latest models from top U.S. developers — including OpenAI and Google — DeepSeek’s R1 is intended to parrot the ways humans sometimes ruminate over problems by spending time computing an answer before responding to user queries. DeepSeek’s version, which is built atop one of the company’s other recently released models, V3, differs from its US peers in its efficiency, however.

The team behind it worked innovatively. While rivals have used a huge number of high-powered computer chips to build similar AI models, DeepSeek team members appear to have found ways to efficiently use the relatively small amount of less-advanced ones they had access to given US export controls on the most cutting-edge chips. And they leaned heavily on a technique known as reinforcement learning that rewards a system for correct answers and punishes it for those that are incorrect.

In the U.S., some tech and policy leaders have acknowledged these advances while also raising questions about whether the Chinese company built its chatbot on the back of Western technology, sidestepping some of the enormous costs of developing large language models, the building blocks of chatbots.

OpenAI said it is reviewing whether DeepSeek “may have inappropriately distilled our models” to build its own rival software. DeepSeek has not responded to comments on the allegation.

Back up. What is distillation?

Distillation refers to using the outputs of a company’s AI to train a different model — typically a smaller, less powerful one — to have similar capabilities. Some companies, such as OpenAI, say it violates their terms of use to use the outputs of their AI models to train a competing model.

What have government officials said about open source?

After conducting a review, in 2024 the administration of former President Joe Biden concluded that it would be premature to impose restrictions on open AI models but also left open the possibility that there may be reason to do so in the future.

The administration of President Donald Trump has yet to fully clarify its policy on artificial intelligence, but some close to the president — including Elon Musk and Vice President JD Vance — have previously expressed strong support for open source AI software.

In the wake of the commotion over DeepSeek, however, White House AI czar David Sacks suggested that private sector developers might make an effort to protect their models from being distilled.

©2025 Bloomberg L.P.