The AI race isn’t just between companies, but between two philosophies: keep powerful models closed, or release them for anyone to use. Here’s what open-source AI is, and why the debate matters.
Artificial intelligence models are trained on huge amounts of data. The result of that training is a set of billions of numbers, called weights, that determine how the model behaves.
Companies take two broad approaches to releasing them.
Closed and open
Closed models keep the weights secret. Users can access them only through the company’s own apps or paid services. Many of the best-known AI models, including those behind ChatGPT, Google’s Gemini and Anthropic’s Claude, are offered this way.
Open models are released for anyone to download. Developers can run them on their own computers or servers, inspect them, and modify or retrain them for their own purposes, usually for free.
Strictly speaking, most are “open-weight” rather than fully open source, because companies usually release the trained model but not all the training data or code used to build it.
The major open models
Meta’s Llama models were among the first powerful open-weight models, and have been downloaded hundreds of millions of times.
In January 2025, Chinese company DeepSeek released a model that performed close to leading U.S. models, which it said it had built for a fraction of the cost. The announcement caused a sharp sell-off in technology shares, with Nvidia losing almost US$600 billion in market value in a single day, as investors questioned how much companies would need to spend on AI.
Other Chinese companies, including Alibaba with its Qwen models, French company Mistral, and U.S. companies including Google and OpenAI have also released open-weight models.
The case for open AI
Supporters argue open models spread the benefits of AI more widely. Researchers, start-ups, governments and businesses can use and adapt them without paying a large company or sending sensitive data to it. That can encourage innovation and competition.
Being able to inspect models also helps researchers understand how they work and find problems.
The case for caution
Critics argue that once a powerful model is released, it can’t be recalled. Safety measures built into it can be removed, and bad actors could use it to create misinformation, scams, malicious software or, in the future, more dangerous material.
There are also concerns about national security. U.S. officials have raised questions about the rapid progress of Chinese open models, and some governments, including Australia’s, have banned DeepSeek’s app on government devices over data security concerns.
The debate is likely to intensify as AI models become more capable.
