Meta opens advanced AI weights to counter Chinese rivals, Zuckerberg pushes U.S. policy shift

Meta announced the launch of Muse Glimmer on Monday, a family of open-source AI models engineered to run on a laptop, marking an intensified commitment to open-source development in the face of both competitive and geopolitical pressures. Alongside this release, the company pledged to open the weights of Muse Spark 1.2, its most advanced model available on the market. Opening the weights means Meta will publicly release the underlying trained parameters, allowing anyone to download, inspect, and modify the model rather than restricting access through Meta's own products or API.

Mark Zuckerberg accompanied the launch with a 6,500-word essay making the case for American open-source AI dominance. In that essay, he argued that powerful AI should not be controlled by a handful of companies—an implicit critique of closed-source competitors like OpenAI and Anthropic. Zuckerberg positioned Meta's open-source commitment as critical to challenging Chinese open-source models from DeepSeek and Moonshot, which he described as getting "uncomfortably close to the American frontier." He wrote: "I do not believe restricting access to foreign open source models is an effective solution. Our goal should be for American open source models to be the best globally. This requires removing the hurdles that make it harder for American open source models to compete." Zuckerberg also urged Washington to support American efforts and argued that American AI developers face disadvantages against Chinese rivals, including in training data.

This announcement is part of a coordinated media strategy by Meta that has included a Zuckerberg-penned opinion piece in the Wall Street Journal on the same themes published last month, along with interviews with select media outlets. The timing reflects both competitive and strategic calculations. Meta initially differentiated itself through open-source when it launched the Llama family in February 2023, which helped the company build influence across the AI ecosystem even as competitors like OpenAI and Anthropic captured more revenue by keeping their models proprietary. Over the past year, however, Meta's commitment to open-source has become less consistent, mixing open releases with proprietary ones. The current pivot back to openness appears designed to regain momentum after recent Meta models have fallen short of systems from OpenAI and Anthropic, and to reassure investors of the company's competitive position.

Meta trained the Glimmer series using distillation, a technique in which smaller models learn from Muse Spark, Meta's larger proprietary system. The company is positioning the Muse family as foundational infrastructure for a more capable Meta AI assistant that will run across Facebook, Instagram, WhatsApp, and its Ray-Ban Meta AI glasses, while also competing directly with models from Anthropic and OpenAI. By coupling product releases with policy advocacy, Meta is attempting to establish both technical and political differentiation in an increasingly contested AI landscape.