Meta releases open-weight Glimmer model for local superintelligence

Thirty billion parameters represent Meta’s latest attempt to define the boundary between public tools and proprietary corporate assets. On Monday, the company released Muse Glimmer, an open-weight model designed to execute multi-step tasks—including code debugging and file management—locally on consumer-grade GPUs. By utilizing an Apache 2.0 license, the release provides a functional template for what CEO Mark Zuckerberg describes as personal superintelligence, operating entirely within the perimeter of a user’s own hardware.

The technical shift toward local execution addresses a fundamental tension in the deployment of personal assistants: the requirement for deep data access. By processing schedules, messages, and sensitive files on-device rather than in the cloud, Glimmer establishes a privacy framework for agents intended to be always-on. The model is trained across more than 100 languages and supports both text and image inputs, positioning it as a localized sibling to the company’s more powerful, closed-weight Muse Spark model.

This release clarifies a strategic bifurcation at Meta. While the company continues to reserve its most advanced capabilities for controlled, closed systems, Glimmer serves as a distribution vehicle for a capable, offline-ready tier of intelligence. In a letter accompanying the release, Zuckerberg argued for the broad distribution of such capabilities as a means of individual empowerment, suggesting that the future of the industry lies in a balance between centralized supercomputing and ubiquitous, locally-owned agents.