Meta open AI strategy centres on smaller, locally run models

Meta’s open AI strategy concerns models it describes as open, rather than a disclosed investment in OpenAI. The company is promoting a new open-weight model that can run on a laptop, as it seeks to catch up with advanced rivals in the West and China, The New York Times reported on August 10.

Mark Zuckerberg has presented the approach as a challenge to the economics of the largest frontier systems. According to the Times, Meta’s aim is to show that open models can be far smaller and cheaper than leading systems. A model capable of running locally on a laptop could potentially widen the range of settings in which artificial intelligence is used, although Meta’s announcement as described by the Times does not set out the commercial or technical terms for doing so.

Why is Meta pursuing open AI models?

Zuckerberg set out Meta’s case in a 14-page essay focused on broad access to AI and AI agents, according to the Times. The examples cited included personalised coaches for different subjects and tools to help people start businesses.

The push continues a strategy evident in Meta’s 2024 release of Llama 3. Meta said then that it wanted to create open models comparable with the strongest proprietary alternatives available at the time, NBC News reported. Zuckerberg also said on a first-quarter earnings call that outside open-source contributions would help Meta’s models improve.

For Meta, a broader developer base could have strategic value while the company competes for technical capability. Alex Ratner, chief executive of Snorkel AI and a computer-science professor at the University of Washington, told NBC News that companies may favour models that allow them to retain ownership of their own data and the resulting output. He said this could make Meta’s technology more attractive to developers and data scientists, and could help it recruit talent and develop revenue opportunities. That was Ratner’s analysis, rather than a stated Meta plan.

Meta also has a record of backing widely available software projects, including PyTorch, React and GraphQL, NBC News noted.

Is Meta’s Llama actually open source?

The answer is qualified for Llama 3. NBC News reported that organisations using Llama 3 in products with more than 700m monthly active users had to seek a licence from Meta. It also reported that Meta had not disclosed the training data for Llama 2 and Llama 3, making it difficult to assess possible bias.

In addition, the Open Source Initiative had not approved Llama 3’s licence, according to NBC News. Those restrictions and omissions have made its open-source label contested in parts of the technology sector. The available reporting does not establish whether the newly announced 2026 model carries comparable terms.

The Times separately reported that Meta is creating a $1bn fund for police and firefighters in communities hosting its data centres, an effort it described as addressing local opposition in states including New York and Texas. That initiative concerns data-centre community relations, rather than the rationale for the company’s AI-model strategy.

This story draws on original reporting from NYT DealBook.