Meta’s New “Runs on Your Mac” AI Needs Two of Apple’s Priciest Desktops to Work

Meta put the weights for Muse Spark Open out on Thursday, building on the Muse Spark model it had kept proprietary since April. 

Meta says the model was built around Apple’s MLX framework and tuned specifically for Mac Studio machines carrying the new M5 Ultra chip. 

Engadget flagged that the setup needs at least a terabyte of RAM to run, well past anything resembling a standard desktop. 

So, getting Muse Spark Open running locally means pairing two M5 Ultra Mac Studios together, not buying one.

A Model Built for a Setup Almost Nobody Owns

A single top-end Mac Studio already costs over eighteen thousand dollars before accessories. Doubling that hardware just to load this model pushes home use out of reach. 

Meta frames this as bringing “leading agentic capabilities to your desktop for the first time,” language that technically holds up but glosses over just how narrow that desktop audience really is. 

For everyone else, Meta points to API access through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI, a route that avoids the hardware entirely.

As the company says, downloadable weights are available on Hugging Face, Amazon SageMaker, Google Cloud Model Garden, and Microsoft Foundry.

Best in the US, Not Best Overall

Meta calling Muse Spark Open the top US-made open-weight model is technically accurate, but it hides a narrower reality. 

On Terminal-Bench 2.1, a coding benchmark, Meta’s own published numbers show Muse Spark Open scoring 76.4, behind Alibaba’s Qwen3.8 and GLM 5.3, and well behind Moonshot’s Kimi K3. 

Things change on MCP-Atlas, which tests tool selection for plain-language prompts. Muse Spark Open scored 89.2, outperforming every competitor on Meta’s benchmark table, Chinese models included.

So the model genuinely leads somewhere, just not in the category most developers associate with a coding-focused release.

What This Launch Actually Signals

Put together, the picture is less “Meta cracked open-weight AI for everyday Mac users” and more “Meta built a serious tool for people who already have serious hardware, and a serious API bill for everyone else.” 

Zuckerberg continues to argue that advanced AI shouldn’t be locked behind corporate doors, and Meta’s previous open-weight model, Muse Glimmer, was indeed a laptop-ready release. But running Muse Spark Open locally is out of reach for most. 

Tying it heavily to Apple’s chip stack and a restrictive license means it is aimed squarely at enterprise teams who won’t hesitate to buy a second Mac Studio.

Source: The Most Capable US Open‑Weight Model to Date