Muse Spark Open Needs 1TB of Memory and Four Mac Studios

Meta has released Muse Spark Open, an 800-billion-parameter model it says is built for local deployment on Apple M5 Ultra hardware. Running it takes 1 TB of unified memory — a figure most buyers can only reach by clustering Mac Studios.

What Meta is claiming

The company describes Muse Spark Open as "the most capable artificial intelligence model in the US to date, among those whose files defining its operation are available for download." That is a claim about downloadable models specifically, not about frontier systems in general.

On performance, the reporting is explicit about provenance: "In tests published by the company, Muse Spark Open outperforms GLM 5.3 and Kimi K3 on tasks that require it to complete complex steps and draw conclusions from images. These are Meta test results, not an independent comparison."

Treat the leaderboard framing accordingly. No third-party reproduction has been published, and the comparison set was chosen by the vendor.

The hardware bill

This is where the story gets concrete for anyone pricing a local deployment. The model "requires 1 TB of unified memory, achievable by clustering multiple Mac Studio units." Specifically, it "can also run on four Mac Studios with 256GB of memory each."

The quoted cost of that cluster: about $38,000.

That figure is for the machines only. It excludes networking between units, power, and the engineering time to shard an 800B model across four boxes — none of which the source accounts for.

There is also a timing wrinkle. The report notes that "Apple hasn't announced a specific price or release date for the 512GB Mac Studio," saying only that it will be available in late October. Anyone planning a two-box configuration instead of four is waiting on pricing Apple has not published.

Who this is actually for

Meta's own positioning is narrow. The model "may be of interest to large organizations that spend a lot on using AI models, as long as it performs as it claims" — a conditional the company itself attaches.

That is the right way to read a $38,000 entry ticket. The arithmetic only works if you are currently spending enough on API inference that a one-time hardware purchase amortises inside a reasonable window, and if the quality holds on your workload rather than on Meta's benchmark selection.

For individual developers and small teams, this is not a buying decision. It is a signal about where unified-memory Macs are being positioned: as inference hardware for models that will not fit on a single GPU.

What to check before budgeting

The weights are already distributed — Muse Spark Open "is available for download from Hugging Face" — so independent evaluation is possible rather than theoretical.

Three things worth waiting for: a third-party benchmark run against GLM 5.3 and Kimi K3 on the same tasks, real throughput numbers from a four-unit Mac Studio cluster rather than a memory-capacity claim, and Apple's pricing for the 512GB Mac Studio, which decides whether the cheapest viable configuration is two machines or four.

Until those land, the only verified facts here are the parameter count, the memory requirement, and the price of the hardware needed to meet it.