Meta announced Muse Glimmer, a new family of open-source / open-weight models specifically designed to run locally on a laptop. It marks Meta’s return to open releases after launching the closed-source Muse Spark in April.
The first release is a 30-billion-parameter model optimized for always-on local agent workflows, small enough to run on a Mac or PC with a single consumer GPU. Meta positions it as lower-cost, customizable, and privacy-preserving alternative to frontier closed models, and as a U.S. response to leading Chinese open-weight models from Moonshot, Alibaba, and DeepSeek.
Details: Muse Glimmer
What it is:
- Name: Muse Glimmer
- Type: Open-weight / open-source LLM, with publicly accessible core components for download and modification
- Size: 30-billion-parameter model
- Family: First model in new Glimmer family, with more models planned soon
Designed to run on a laptop:
- Small enough to run on a Mac or PC with a single consumer GPU / single graphics card and
- No cloud required, enabling offline, local execution
- Optimized for consumer hardware to reduce cost and latency
What it’s built for:
- Agentic tasks: Designed for always-on local agent workflows
- Use cases listed by Meta: local agents and function calling, local coding, and LLM-as-a-judge evaluation
- Built with a novel distillation recipe and fine-tuning for reliable tool use and multi-step reasoning
Strategic Context:
- Meta was an early champion of open-source AI, then shifted to closed model Muse Spark in April after rebuilding its Superintelligence Labs
- Meta said it plans in the coming weeks to open the weights for a version of Muse Spark 1.2 as well
- Zuckerberg statement: U.S. needs lower barriers for open-source models to compete, and “Rather than centralizing superintelligence, we should distribute it widely”
- CEO Mark Zuckerberg also advocated for model distillation – using a powerful system to train smaller models – and said Meta will add governance for safety criteria
- Launch comes as open-weight models are seen as cheaper and more customizable for enterprises vs. closed models from OpenAI and Anthropic
