Meta Open-Sources Muse Spark 1.2, Announces On-Device Model Family Muse Glimmer

Meta will open-source the weights of its latest AI model, Muse Spark 1.2, and launch a new family of laptop-optimized models called Muse Glimmer, CEO Mark Zuckerberg announced Monday in an Instagram video and an accompanying 6,500-word essay.

Opening the weights means the public can download and run Muse Spark 1.2 directly, rather than accessing it only through Meta's hosted API. Muse Glimmer, according to Meta, is designed to run locally on laptop hardware rather than in cloud data centers — a departure from the compute-heavy inference model used by most frontier AI products today.

Meta did not disclose parameter counts, context window sizes, benchmark scores, or pricing for either model. Specific technical documentation has not yet been published.

The competitive positioning

Zuckerberg framed the release as a direct challenge to both closed-model U.S. labs and open-weight Chinese labs. OpenAI and Anthropic have largely kept their most capable models closed, while Chinese companies including Alibaba, DeepSeek, and Moonshot AI have released open-weight models that compete with U.S. offerings on several benchmarks.

"Foreign labs currently hold several advantages here since American labs have to comply with many additional restrictions on training data," Zuckerberg wrote. "US policy must reduce this additional friction if we want American open source models to lead over time." He called for policy changes around data use and distillation — the practice of training new models on the outputs of existing ones — arguing restrictive U.S. rules put American labs at a disadvantage rather than protecting them.

Industry analysts see the on-device push as a distinct strategic bet. "Bringing small, agentic models like Muse Glimmer directly onto PC and mobile hardware bypasses cloud compute costs to outcompete Google, Microsoft and others on the end-user's device," Neil Shah, co-founder of Counterpoint Research, told CNBC.

Investor context

The announcement comes as Meta faces investor pressure over AI spending. The company has forecast capital expenditures of up to $145 billion this year, largely tied to AI infrastructure through Meta Superintelligence Labs, formed in 2024. Meta shares are down roughly 10% year-to-date but rose 2.1% in premarket trading Monday following the announcement.

Zuckerberg also used the essay to distance Meta from safety rhetoric common among rival labs. "It is surprising that the discourse from many developing AI is so filled with doom," he wrote, adding that "the notion that AI is so dangerous that the only safe path is an extreme concentration of power seems inherently problematic" — a comment widely read as directed at Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman, both of whom have warned publicly about AI's impact on jobs.

What this means

Meta is betting that open weights, not closed APIs, will win developer mindshare as Chinese labs flood the market with competitive open models. Muse Glimmer's on-device design targets a real gap — most frontier models still require cloud inference — but without published benchmarks, context windows, or pricing, it's impossible to verify whether Muse Spark 1.2 or Glimmer actually compete with closed models from OpenAI and Anthropic on capability. The essay's policy arguments matter as much as the models themselves: Zuckerberg is lobbying Washington to loosen data and distillation restrictions specifically to help U.S. open-source labs, Meta chief among them, close the gap with Chinese competitors.