Skip to content
NewsModel release

Meta ships an 816-billion-parameter open model built to run on two Mac Studios

· by Pondero Newsdesk

The short version

Meta's October 15 release of Muse Spark Open pairs an 816-billion-parameter, 1-million-token-context model with a license letting it run entirely on two Apple M5 Ultra Mac Studios instead of a cloud GPU cluster.

Meta ships an 816-billion-parameter open model built to run on two Mac Studios

Meta released Muse Spark Open on October 15, an open-weight model with 816 billion total parameters, 42 billion of them active per token, and a 1-million-token context window, and the company built it specifically to run on a pair of Apple Mac Studios instead of a rented GPU cluster per Meta's research blog. That design choice is the actual story: a model this large has never shipped with instructions for running it entirely on desktop hardware a team can own outright.

What

Muse Spark Open is natively multimodal and tuned for coding and agentic work, including long-horizon tasks and multi-step reasoning, per Meta. "The most capable US open-weight model to date" is how the company bills it, a claim TechCrunch pointed out rests on evaluation results Meta published itself rather than independent testing per TechCrunch. It ships under Meta's own Muse Model License, which the company says carries "limited commercial terms" that let many developers use it for free. Access is already wide: Together AI, Fireworks, Baseten, Modal, Groq, Nebius, and OpenRouter all host it, downloadable weights sit on Hugging Face, Google Cloud Model Garden, Amazon SageMaker, and Microsoft Foundry Models, and it runs on inference engines including vLLM, TensorRT-LLM, SGLang, Apple's MLX, and llama.cpp [per Meta].

On the inference-engineering side, Meta says the open-weight build ships with custom, auto-tunable Helion GPU kernels inside SGLang that deliver a 1.03x to 3.95x kernel-level speedup on Nvidia and AMD GPUs, plus a 3% end-to-end speedup on AMD's MI350 per Meta's research blog. That is a narrower, more technical claim than the headline parameter count, and it signals Meta wants this model taken seriously on non-Apple hardware too, not just as a Mac Studio showcase.

Why Apple Silicon, specifically

Meta built the local path around one chip generation: the M5 Ultra. Apple's unified memory architecture lets that chip's integrated GPU address up to 512GB per Mac Studio, and its memory bandwidth runs at 1.2 terabytes per second, fast enough, Meta says, to read the small slice of weights each token needs without stalling per Meta. Chain two Mac Studios together and the combined unified memory crosses the one-terabyte mark Meta calculates the full model needs, with the pair behaving, in the company's words, "like one supercomputer." That is a specific hardware bet, not a generic "runs on consumer GPUs" claim: Meta picked the one widely available desktop platform that could plausibly hold an 816-billion-parameter model's weights without a rack of data-center accelerators.

Meta also published evidence the setup does real work, not demo-only tasks. Its Muse Code agent, running on Muse Spark Open entirely on-device, shipped bug fixes, new features, and a full build to Ice Cubes, an open-source Mastodon client, and in a separate test drove DaVinci Resolve through an end-to-end editing pass, reviewing footage frame by frame to find the strongest moments and crop them for vertical video per Meta's research blog. Those two examples point at the same underlying case: coding agents and creative-tool agents that run long, multi-step sessions benefit from execution that never leaves the machine, where there is no per-token metering ticking and no round trip to a vendor's API for every tool call.

Running it locally still means buying two top-end Mac Studios

That local-first framing matters most for teams that need a frontier-scale coding model to stay on hardware they control rather than inside a vendor's cloud, whether for contractual data-residency requirements, a fixed compute budget instead of metered API spend, or simply distrust of sending proprietary codebases to a third party. Muse Spark Open gives that group an option that did not exist at this parameter count: buy the hardware once, and every subsequent coding session runs without a network call to Meta, Together AI, or any of the six other hosted platforms.

None of that comes cheap up front. Two Mac Studios configured with M5 Ultra and enough unified memory to clear Meta's own threshold are a serious capital purchase, and nothing about the local path beats the per-token price of a hosted API call once a team amortizes the hardware cost against light or sporadic usage. The tradeoff is upfront hardware spend and setup time against an ongoing, metered bill, and the two Mac Studio configuration only wins for teams running the model heavily enough, or with a compliance requirement strict enough, to justify owning the box rather than renting the inference.

Context and reactions

Meta's own benchmark page compares Muse Spark Open against a specific peer set: Inkling, Nemotron 3 Ultra, Beam, Mistral 4 Large, Qwen3.8-2.4T-A95B, GLM 5.3, and Kimi K3, scored across agentic, knowledge-and-reasoning, coding, and tool-assisted multimodal evaluations per Meta's research blog. The scores themselves appear only as an image on Meta's page rather than in machine-readable text, so Pondero is not reporting specific numbers from that chart. Meta's chosen comparison set does not include DeepSeek V4, one of the open-weight models most frequently discussed in the same frontier-scale, coding-focused class, and it compares against Qwen3.8-2.4T-A95B rather than a newer Qwen3.6 release. Both omissions matter for anyone trying to map Muse Spark Open's standing against the specific competitors they already use.

TechCrunch framed the release inside the broader open-weight argument: models with published weights can be more cost-efficient to run at scale and offer transparency and flexibility that closed, proprietary models do not, benefits TechCrunch says apply generally to the category rather than specifically to this release per TechCrunch. Meta frames its own motive for releasing weights at all in similar terms, writing that open releases are "part of how we bring the benefits of personal superintelligence to everyone, and part of how safety work advances across the field," and the company says it ran Muse Spark Open through its Meta Superintelligence Scaling Framework and that the model met the thresholds that framework requires for release, with a full evaluation report promised "available soon" alongside the model card already live on Hugging Face per Meta's research blog.

What to watch next

The clearest open question is independent verification: whether outside researchers replicate Meta's comparison against Kimi K3, GLM 5.3, and the rest of its self-selected peer set, and separately, whether anyone runs Muse Spark Open against DeepSeek V4 or a newer Qwen release, the two obvious omissions from Meta's own chart. Also worth tracking is whether other inference vendors beyond the seven Meta already lists add native M5 Ultra optimizations now that a frontier-scale model is tuned specifically for that chip, and whether Meta follows through on the promised full safety evaluation report.

Sources