Ever wanted a frontier-scale AI living under your desk? Good news: it’s now a thing. Slightly less good news: your desk needs two very expensive Macs. On October 15, Meta released the weights for Muse Spark Open, an 816 billion parameter model it calls “the most capable US open-weight model to date.” It teamed up with Apple so the whole thing runs locally on a pair of M5 Ultra Mac Studios.
What’s in the box
Per Meta’s announcement, Muse Spark Open is a natively multimodal mixture-of-experts model with 816 billion total parameters, 42 billion active for any given token, plus a 1 million token context window. Translation: it’s huge, but it only wakes up the parts it needs. It’s built for coding and agent work, meaning long, multi-step jobs rather than “write me a haiku about my cat.”
You can grab the weights on Hugging Face, Amazon SageMaker, Google Cloud and Microsoft Foundry. Or skip the hardware and use hosted versions from Together AI, Fireworks, Groq, Baseten, Modal, Nebius and OpenRouter. It ships under the new Muse Model License, which Meta says lets many developers use it free. Bigger companies serving it as a commercial AI service need a separate agreement.
The Apple trick
This is the fun part. Apple’s unified memory lets the M5 Ultra’s GPU use up to 512GB per Mac Studio, so two machines linked over Thunderbolt 5 can hold the entire model. Meta says the chip’s 1.2 TB/s memory bandwidth is quick enough because each token only needs a small slice. In Meta’s words, two Mac Studios “behave like one supercomputer with a terabyte of unified memory.” That’s a sentence that would have sounded bonkers a few years ago.
Meta says it ran real agent workflows entirely on-device, like shipping bug fixes and features to an open-source iOS app, and driving DaVinci Resolve to pick and crop the best moments from video footage. 9to5Mac covered the launch, and Mark Zuckerberg and Apple CEO John Ternus even traded public thank-yous. Before you start clearing desk space: the 512GB Mac Studio isn’t on sale yet. Apple has promised it for late October, per reporting relayed by AI News Online. And even the 256GB model costs over $10,000. So, yes, about that wallet.
How does it stack up?
On Meta’s own model card, Muse Spark Open leads open rivals on agent benchmarks like MCP-Atlas (89.2). It still trails China’s top open models, Kimi K3 and GLM 5.3, on some coding tests and on the Artificial Analysis index. Those are vendor numbers, so salt to taste. Still, this is the closest a US open-weight model has come to the Chinese open leaders in a while. Meta is keeping its very best model, Muse Wonder, behind closed doors.
Desk AI is having a moment
Same day, Nvidia dropped its open Nemotron 3.5 Super VL agent model as RTX Spark laptops from Microsoft, Asus, Dell, HP, Lenovo and MSI hit shelves. Running serious AI on your own machine, without shipping your data off to somebody’s data center, went from nerdy hobby to real product category in a matter of months.
Why it matters for EVs and robots
Local, open, frontier-class models are a big deal for anyone building machines that can’t always phone home, from robotics labs to automotive software teams. Being able to inspect, fine-tune and run a top model on your own hardware lowers the bar for building smarter robots and smarter cars. Cars and humanoids will still run their own purpose-built AI chips, of course, but tools like this shrink the gap between a garage startup and a giant.
The EVX News take
Two Mac Studios won’t turn your home office into a data center, but they get surprisingly close. Open weights plus beefy consumer hardware is a healthy counterweight to an industry that keeps its best toys locked behind APIs. Now we’re wondering if a “home AI server” will one day sit next to the Powerwall in the garage. We’d buy that.
Sources: Meta AI Research, Hugging Face, 9to5Mac, AI News Online.
— The EVX News team
Featured image: Meta, from the Muse Spark Open announcement.
