Meta has announced a renewed push into open-weight AI models, releasing a new model called Muse Glimmer and pledging to publish the weights of its more capable Muse Spark 1.2 model within weeks. The move, accompanied by a lengthy essay from CEO Mark Zuckerberg on AI philosophy and governance, is Meta's latest attempt to reposition itself against rivals such as OpenAI and Anthropic, whose proprietary approach has drawn lobbying efforts against distillation and Chinese open-weight competitors. It marks another shift in strategy for Meta's AI division after a turbulent year of internal reshuffles and repeated changes of direction.

Muse Glimmer is a 30-billion-parameter model with a 128,000-token context window, distilled from the larger Muse Spark and released under the open-source Apache 2.0 licence for local use on consumer machines. This follows Muse Spark's launch in April as a closed, proprietary system, and July's Muse Spark 1.1 update, which introduced Meta's first paid AI service. Muse Spark 1.2, released on 5 August alongside a coding agent called Muse Code, has been seen by developers as cost-competitive but not matching frontier rivals in capability. Zuckerberg's essay defends distillation as a legitimate practice and champions decentralised, widely distributed AI, echoing a joint industry letter signed in July by Meta, Nvidia, Hugging Face, Mistral, OpenAI and others in support of open-weight models.