Meta built a frontier open-weight model and pointed it at your Mac, not NVIDIA's cloud

Meta released Muse Spark Open on October 15 — what it’s calling the most capable US open-weight model to date, built for coding and multi-step agentic work. The headline numbers are real (816 billion total parameters, 42 billion active via a mixture-of-experts design, a 1 million token context window), but the number worth sitting with isn’t in the benchmark table. It’s who Meta built this to run on: not NVIDIA’s data-center GPUs, but Apple Silicon Mac Studios, via Apple’s own MLX framework.

What’s actually happening

Muse Spark Open ships with support across the usual inference stack — SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, llama.cpp, Unsloth — so it runs on NVIDIA and AMD hardware just fine, and Meta even built custom “Helion” GPU kernels that squeeze out a 1.03–3.95x speedup there. But Apple MLX gets named specifically and first in Meta’s own materials, with on-device demos (an iOS client build, a DaVinci Resolve video-editing workflow) run entirely on Mac Studio hardware. Weights are open on Hugging Face, Amazon SageMaker, Google Cloud, and Microsoft Foundry; API access exists too, through Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter, and Together.

Why the Apple-Silicon framing is the real story

Meta is, by a wide margin, one of NVIDIA’s biggest customers — its own data centers run on tens of billions of dollars of NVIDIA training silicon. That makes it notable, not incidental, that when Meta wanted to show off this model running privately on someone’s own hardware, it reached for Apple Silicon instead of its own GPU fleet. That’s the same local-inference bet I’ve been tracking all year from a different direction — NVIDIA’s PAIR tool clustering your own Macs and RTX PCs, NVIDIA and Microsoft’s agentic-laptop push, Apple’s “no cost per token” pitch to enterprise buyers. Each of those came from a hardware vendor with obvious skin in the game. This is a model-layer lab — one of the handful of companies capable of training something this size — independently pointing at the exact same destination: AI capable enough to matter, running on hardware you own and keep in your own building. When the people selling chips and the people training frontier models converge on the same answer from opposite ends of the stack, that’s a stronger signal than either one alone.

The honest caveat

“Runs on a Mac” undersells the actual ask here. Meta’s own on-device configuration is a pair of maxed-out M5 Ultra Mac Studios, networked together, with up to a terabyte of unified memory apiece. That’s serious, enthusiast-grade (or well-funded small-team) hardware — not something most businesses have sitting in a closet, and a real step up from the single-workstation story I’ve told about models that fit on one machine. Worth also flagging before anyone builds a commercial product on this: the weights ship under Meta’s own Muse Model License, not a fully permissive Apache/MIT-style release — it carries real commercial-use terms that need an actual read, not an assumption, before you put it into anything client-facing.

What this means practically

The specific hardware bar will come down — it always has, model-generation over model-generation, the same way last month’s “needs a data center” became this month’s “needs two Mac Studios.” The part that doesn’t depend on waiting for that is the direction: three different kinds of companies — a chip vendor, an OS maker, and now a frontier model lab — are all independently betting that local, on-prem inference is where real capability is headed, not just a privacy-minded compromise on capability. That’s the environment a readiness audit is built to evaluate against: not “can you afford to run this exact model today,” but whether your infrastructure and licensing homework are actually ready for the version of this that does fit your hardware, whenever that lands.