Meta doubles down on open AI as Zuckerberg calls for fewer US restrictions

Meta is stepping up its bet on open artificial intelligence, with Mark Zuckerberg arguing that the United States risks falling behind China unless it makes it easier for companies to develop and release open-weight AI models.

The Facebook owner on Monday unveiled Muse Glimmer, a relatively small AI model designed to run agentic tasks on an ordinary Mac or PC equipped with a single graphics card.

The idea is straightforward: instead of relying entirely on expensive cloud-based AI, powerful models could increasingly run directly on people’s own devices.

Zuckerberg said bigger models are also on the way. In a 14-page essay titled “The Future is for Everyone”, he argued that AI should not become the preserve of a handful of companies.

He also rejected the idea that concentrating AI development among a few powerful firms is necessarily the safest approach.

The argument comes as businesses face growing AI costs and increasing concerns about the security of powerful closed models. Open-weight systems are generally cheaper and allow developers to inspect and customise key parts of the technology.

China has emerged as a major force in this area, with models from companies such as Moonshot, Alibaba and DeepSeek increasingly competing with leading US systems.

Meta plans to release the weights of Muse Spark 1.2, its most advanced model, after previously changing course following disappointing reception for Llama 4.

Zuckerberg also called for changes to US policy covering training data and AI model distillation, saying American developers face restrictions that can put them at a disadvantage to foreign competitors.

Meta is simultaneously trying to address resistance to the huge data centres needed to power AI. The company plans to create a $1 billion fund to support communities affected by its infrastructure expansion.

For Zuckerberg, the message is clear: winning the AI race will require not only better models, but also cheaper access, fewer barriers and enough infrastructure to run them.