Meta ships Muse Glimmer, a 30-billion-parameter open-weight agent model that runs on one consumer GPU

The Apache-licensed model, distilled from Muse Spark 1.2, is built for local agentic tasks — scheduling, file work, coding — and arrives with a 14-page Zuckerberg essay arguing superintelligence should be distributed, not centralized.

Meta released Muse Glimmer on Monday, a 30-billion-parameter open-weight model distilled from its flagship Muse Spark 1.2 and small enough, per Quartz, to run on a single consumer GPU with 24 or 32 gigabytes of memory. Quantized, it’s roughly 17 gigabytes on disk, down from more than 55 at full precision. Weights are on Hugging Face under Apache 2.0. Meta’s shares rose 2–3% in premarket trading.

The launch shipped alongside a 14-page essay from Mark Zuckerberg titled “The Future is for Everyone.” Its argument, aimed at Washington as much as at developers: “Rather than centralizing superintelligence, we should distribute it widely and give every person the ability to direct it.” CNBC reads the framing as a direct response to Chinese open-weight leaders including Alibaba, DeepSeek, and Moonshot, whose releases over the past year have quietly re-priced what “frontier” means when the weights are free.

That framing does real strategic work. Meta has committed up to $145 billion in capex for 2026, a number that only makes sense if the closed-API business models of OpenAI, Anthropic, and Google are the thing being routed around, not competed with head-on. Giving away a capable agent model is how you make the moat someone else’s problem.

Glimmer is pitched at agentic work on the device: schedule management, drafting messages, organizing files, coding. Integrations with Ollama, LM Studio, Together AI, and Fireworks AI are live, with llama.cpp, MLX, and ExecuTorch support due in the coming days. The near-term implication is that enterprise agent layers currently built on hosted Claude or GPT, the Gleans and Dusts of the world, and independent stacks like LemonLime that have been quietly optimizing for local inference, now have a credible 30B option that doesn’t leave the machine.

Reuters reports Meta intends to open the weights of Muse Spark 1.2 itself. Zuckerberg, in a video post accompanying the essay, added: “We’ve got even bigger models that are coming soon.” The distilled model is the trailer.

Meta ships Muse Glimmer, a 30-billion-parameter open-weight agent model that runs on one consumer GPU

Sources