Meta releases Muse Glimmer, Zuckerberg defends distillation
Meta is back in the open-weights game, and Mark Zuckerberg is picking a fight with the closed-model labs while he's at it.
Meta opened the weights for Muse Glimmer, a 30B-parameter dense model the company calls one of the highest-performing of its size — small enough to run on a laptop or a single consumer GPU. The release, announced Monday alongside a post from Zuckerberg, ends a four-month pause in Meta's open-model output: since April, the Muse lineup built by Meta Superintelligence Labs has centered on Muse Spark, a closed flagship that marked a quiet break from the Llama-era open playbook. Zuckerberg says the weights for Muse Spark 1.2, Meta's latest foundation model, will follow "in the coming weeks" — which would put Meta's frontier-tier model back in open hands just as the field shifts toward small, locally runnable models and agentic systems.
The timing is not innocent. In an interview with the Financial Times, Zuckerberg publicly defended distillation — training a model on another model's outputs — as a principle, and criticized "closed" AI model makers for espousing a discourse he says is "so filled with doom." That is a direct counter to the criticism Meta absorbed in April, when Muse Spark was reported to have been trained partly via distillation from third-party open models like Qwen; rivals and safety researchers seized on the practice as a shortcut. Zuckerberg's argument is that open weights are the antidote to the concentration his critics fear: in a separate essay outlining a "positive AI philosophy" centered on individual empowerment, he predicted personal superintelligence could increase employment, reviving the vision he first sketched in his 2025 manifesto.
Meta also previewed a $1 billion "Future is for Everyone" fund to invest in US communities near its data centers — an acknowledgement, in effect, that the buildout's costs land on neighbors while its benefits flow to shareholders. It is the rare open-weights announcement bundled with a local politics play, and it suggests Meta wants the goodwill that open releases generate as much as the developer mindshare.
The real story is the straddle: Meta wants closed frontier models feeding its products and API business, open weights keeping the ecosystem's gravity, and distillation defended as the legitimate bridge between the two. Whether the position holds depends on what Spark 1.2's weights look like when they land — and on whether the open community forgives the four months it spent wondering if Meta had gone the way of its rivals.
What to watch: Muse Spark 1.2's open release and its benchmark scores — the first real test of whether Meta's open-weights comeback has substance behind the rhetoric.
Distillation is standard practice across the industry — is it copying, or competition? Tell us in the comments.
