Meta Releases a Free AI Program as Powerful as the Ones Companies Charge For

Llama 3.1, in its largest version, matches OpenAI's and Anthropic's best on standard tests, and anyone can download it. Mark Zuckerberg says open AI is 'the path forward' and compares it to Linux.

Meta released Llama 3.1 today, and its largest version — 405 billion parameters, trained on 16,000 of Nvidia's most powerful chips — is by the company's testing roughly as capable as GPT-4o and Claude 3.5 Sonnet. Unlike them, it is a file. Anyone can download it, run it on their own computers, modify it, and build a business on it, at no charge.

Mark Zuckerberg published an essay with the release arguing that open models will win, the way the open-source Linux operating system won over the closed systems it competed with in the 1990s. Developers don't want to depend on one company's program, he wrote; countries don't want to depend on another country's; and Meta doesn't sell access to AI, so giving the program away costs it nothing and makes everything built on it easier for Meta to use. He also argued, against the safety objection, that open systems are safer because more people can examine them.

The gap between the best AI you can buy and the best you can have for free has been closing since the first Llama leaked eighteen months ago. As of today, by Meta's numbers, it is roughly zero, with the closed labs a few months ahead at most. That has consequences beyond price. It means that whatever the frontier labs can do, everyone will be able to do soon after — governments, startups, and people with worse intentions alike.

Meta is spending tens of billions on the chips and data centers to keep this up. It expects to make the money back on what people build, and on being the company whose AI the world runs.