Yesterday, Meta did something that would have been unthinkable five years ago: it gave away the weights to a frontier-scale AI model. Muse Spark Open is an 816-billion-parameter mixture-of-experts model that Meta calls “the most capable U.S. open-weight model to date,” and — here’s the part that made me sit up — it was designed to run locally on Mac hardware.

The specs, as Meta reports them

Because you should know whose numbers these are: 816 billion total parameters with only 42 billion active per token (that’s the mixture-of-experts trick — most of the model sleeps while a small slice works), a 1-million-token context window, and native multimodality: it reads text, images, audio, and video, and writes text back. It’s built for coding and agentic work — long multi-step tasks, tool use, fixing real software bugs.

Meta’s own benchmarks show it beating two leading Chinese open models, GLM 5.3 and Kimi K3, on agentic tasks and visual reasoning — take that as a vendor claim, not gospel. To demo it, Meta paired the model with Muse Code, its AI coding terminal: it patched bugs in a Mastodon client (“bug fixes, new features, full build”) and drove DaVinci Resolve to review and crop video footage. Ambitious demos, but they do show what “agentic coding” means in practice: the model working inside your tools, not just chatting.

The Meta–Apple twist

Here’s the genuinely interesting hardware story. Meta worked with Apple to optimize the model for Apple’s MLX framework, so it can run across two 512GB M5 Ultra Mac Studios — or four 256GB units — pooled into what Meta calls “one supercomputer with a terabyte of unified memory.” Apple’s unified memory architecture is doing the heavy lifting: the M5 Ultra’s integrated GPU can address up to 512GB per Studio, and the chip’s 1.2 TB/s of memory bandwidth is fast enough to feed the small slice of the model each token needs. The machines link over Thunderbolt 5; in Apple’s tests, four Studios generated over 36 tokens per second.

The compressed weights alone are about 493GB, which is why a single machine can’t hold it. The catch, as ever, is money: the 512GB Mac Studio hasn’t shipped yet (expected later this month), the 256GB version is already over $10,000, and a four-machine setup runs roughly $38,000. This isn’t for vibe coders — it’s for labs and companies that don’t want their code leaving the building.

And yes, Meta CEO Mark Zuckerberg publicly thanked Apple CEO John Ternus, who replied in kind. Two companies that have spent a decade fighting over privacy and advertising are now co-engineering desktop AI hardware. The incentives align: Meta wants its open models everywhere, and Apple wants the Mac Studio to be the canonical machine for serious local AI.

What “open-weight” actually gets you

The weights are free to download right now — Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models — with a license allowing limited commercial use. Day-one API partners include Together AI, Baseten, Fireworks AI, Groq, Modal, Nebius, and OpenRouter, so you can try the model without a terabyte of RAM under your desk.

But note the distinction: open-weight isn’t open-source. You don’t get the training data or the full training code, and the license has teeth — per Particle’s summary, it requires “Built with Muse Spark Open” attribution, and companies over $20 million in revenue must notify Meta and sign a commercial agreement before offering it as a service. Worth reading the actual license before you build anything commercial on it.

If the full model is out of reach, the community has already quantized it down: one developer write-up notes formats as small as ~2-bit that reportedly run on a 256GB machine via llama.cpp — with the usual “quality trade-offs apply” caveat.

Why this matters to the rest of us

The headline isn’t really the parameter count. It’s that frontier-scale models are becoming appliances — things you can download, own, and run on hardware that sits on your desk, tuned for the machines many of us already use. Open-weight models won’t replace hosted APIs for most work, but they change the economics: no per-token bills, no data leaving your network, and the freedom to fine-tune and inspect.

So the question I’ll leave you with: if you had the hardware budget, would you rather rent your AI by the token forever, or buy $38,000 of Mac Studios and own your frontier model outright? I suspect the answer splits neatly between “my homelab would love this” and “the cloud is a perfectly fine someone else’s computer.”