Apple ran Meta's 816B Muse Spark Open on four Mac Studios at over 36 tokens a second
Apple has run Muse Spark Open, an 816-billion-parameter model from Meta with 42 billion parameters active at a time, across four Mac Studios and measured over 36 tokens a second. Apple’s machine learning researchers say the model "cannot fit on a single machine", and their figures put four M5 Ultra machines splitting it by width at 36.73 tokens a second at an 8,192-token context, against 15.74 for two machines in a pipeline.
The banner Apple published with the research post on October 15, alongside the MLX release the benchmarks were run with. Image: Apple
Memory is why it takes four. Muse Spark Open "would require more than 1.6TB of RAM in its native precision (BF16) for the model parameters alone", Apple says, and about 493GB once quantised to 4-bit NVFP4, against a ceiling of 512GB of unified memory in a single Mac Studio.
Apple calls it "the newly released Muse Spark Open", lists it among the "popular open-weight LLMs" it benchmarks, and in its own command lines loads it from a path reading meta-models/Muse-Spark-Open-816B-A42B-NVFP4. The hardware numbers here are Apple’s, measured with its own mlx_lm.benchmark tool.
Thunderbolt 5 RDMA is what holds the cluster together
macOS 26.2 added a low-latency RDMA driver over Thunderbolt 5, and MLX v0.30.1 and later builds on it with a set of communication collectives Apple calls jaccl, "pronounced 'jackal'". Each pair of machines is wired peer to peer with two Thunderbolt cables to double the link bandwidth, and a job goes out across the nodes with one mlx.launch command.
The second cable is worth a number. On four M5 Ultra machines at a 65,536-token context, Apple reports time to first token falling from about 50.5 seconds with one cable per link to 40.6 seconds with two.
Splitting a model two ways is the easy case, and Apple describes pipeline parallelism, which gives one machine the first half of the layers and the other the second, as "extremely simple to implement". Splitting it by width across four machines is the harder one, because every layer has to broadcast its inputs and collect its outputs for each token, which is where the low-latency link earns its place.
The M5 Ultra’s bandwidth is what moves the tokens
Generation speed is bound by memory bandwidth, and the M5 Ultra carries 1.2TB/s of unified memory bandwidth against 819GB/s on the M3 Ultra. Apple reports decode running up to 1.6 times faster on the newer chip and time to first token up to 4.6 times faster, since the prefill is compute-bound and leans on the M5 GPU’s neural accelerators, the subject of an earlier Apple post on that chip.
Power stays domestic. Peak draw is about 200W for a Mac Studio with M3 Ultra and 300W with M5 Ultra, which Apple describes as "a fraction of the power used by a common toaster".
Smaller models need none of this. Qwen 3.8 at 27B takes up to 68GB and Mistral Medium 3.5 at 128B up to 158GB at a 65,536-token context, both inside one 512GB machine, and the same machine runs them at 19.69 and 7.70 tokens a second at the shorter context.
The opposite experiment is three days older. Anthropic’s Frontier Red Team got 0.39 tokens a second out of a 26-billion-parameter model spread across 15 Android TV boxes on October 12, on boxes costing about $55 each with 4GB of RAM apiece, and those are Anthropic’s figures from its own report rather than Apple’s.
Apple’s headline figure comes with its conditions: four Mac Studios with M5 Ultra, the model split by width, an 8,192-token context and 128 generated tokens. At that context, four machines also cut time to first token to 4.8 seconds, against 8.2 seconds on two split the same way.
