Meta Releases 30B-Parameter Muse Glimmer AI Model for Local AI Agents

Meta has made another major move in open AI by releasing the weights of Muse Glimmer, a 30-billion-parameter artificial intelligence model designed to power AI agents directly on consumer computers. Unlike cloud-dependent systems that send requests to remote data centres, Muse Glimmer is designed for local deployment, potentially giving developers greater privacy, offline access and control over how AI processes data.

Meta says Muse Glimmer can run on a single consumer GPU and has been compressed to roughly 4-bit precision, bringing its size below 20GB. The model is designed as an agentic AI system, meaning it can work through multi-step tasks, use external tools, maintain plans, process images, write and debug code, and recover from errors. This could allow developers to create assistants capable of performing tasks on a user's computer rather than merely providing instructions.

The release also reinforces Meta's open-weight AI strategy. CEO Mark Zuckerberg has described Meta as a strong supporter of open source and said the company is preparing to release the weights of Muse Spark 1.2, its latest foundation model. The move comes as Meta continues to position open AI models as an alternative to closed, cloud-based systems developed by companies such as OpenAI, Anthropic and Google.