Meta releases 816-billion-parameter Muse Spark Open model for Mac

By Splitfeed EditorialLast updated Oct 18, 2026, 22:00 UTC

On October 15 and 16, 2026, Meta released Muse Spark Open, an open-weights model with 816 billion parameters, built in collaboration with Apple and optimized for Apple M5 Ultra hardware. To run the model locally, users need two paired M5 Ultra Mac Studio desktops, which together cost around $21,000 and require at least a terabyte of RAM. Meta positioned the release as the most capable open model in the United States, targeting coding and agentic AI tasks, and framed it explicitly as a challenge to competing Chinese open models.

The Meta-Apple pairing is notable because it routes inference through Apple silicon rather than Nvidia GPUs, which have dominated this space. The collaboration signals that consumer-grade desktop hardware is being pushed into territory once reserved for data centers, though the $21,000 price point and specialized hardware requirements mean the audience remains narrow. The licensing terms also complicate the picture: the model is described as open-weights, but its license includes restrictions that stop short of fully open source, raising questions about how "open" the release actually is.

The anti-AI camp is largely dismissive. Chris Person from @papapishu.bsky.social called the release "not for real people" and said there was no reason to treat it as anything other than theoretical, partly because Nvidia has pulled back from consumer hardware. Paul Schöppach at @paulschoppach.org used the announcement as a prompt to advise people to abandon Meta products entirely, and the sentiment was echoed by @bluefreedomquest.bsky.social. A post from @divedoc.bsky.social raised broader concerns about always-listening AI devices from Apple and Meta. The pro-AI camp has published no direct reactions to this release; that camp would typically celebrate a large open-weights model as a win for accessibility and domestic AI competitiveness, though no one has said so here. The middle ground, represented by The New Stack at @thenewstack.io and affine at @refinement.systems, offers a more measured read: the model matches Chinese rivals on agentic benchmarks, but the license restrictions undercut its open-weights claims, and the Mac Studio is becoming a surprisingly mainstream inference platform even at its current price.

The clearest thing to watch is whether Meta clarifies or revises the license terms, since that would determine whether developers can actually build on the model freely. Independent benchmark comparisons against the Chinese models Meta named as rivals would also test whether its performance claims hold up outside of Meta's own evaluations.