The Pulse
Apple Runs Meta’s 816B Muse Spark Open on Four Mac Studios
Apple says its latest MLX release runs Meta’s 816-billion-parameter Muse Spark Open across four M5 Ultra Mac Studios at 36.73 tokens per second. The company’s benchmark details how Thunderbolt RDMA and tensor parallelism let Mac clusters ru
AI.info Team · Written with AI ·
“On a cluster of four Mac Studios, Muse Spark Open generates over 36 tokens per second.”
Apple Machine Learning Research
Four Mac Studios run an 816-billion-parameter model
Apple’s Machine Learning Research team published the benchmark on October 15 alongside a new release of MLX, its open-source machine-learning framework for Apple silicon. The demonstration runs Muse Spark Open, an 816-billion-parameter model, across four Mac Studio systems equipped with M5 Ultra chips. Apple reports 36.73 tokens per second for generated text at an 8,192-token context length.
The result addresses a basic hardware constraint: a model of this size does not fit on one Mac Studio. Apple’s post says a single M5 Ultra can have up to 512GB of unified memory, while its benchmark distributes the model across multiple computers. The company evaluates the model in NVFP4 precision, a detail that matters when interpreting the size and speed claims.
Apple’s figures come from its own benchmark, not an independent comparison. The post measures decode speed after an 8,192-token context and reports results over 128 generated tokens. For Muse Spark Open, Apple does not provide a single-Mac baseline because the model does not fit on one machine.
Tensor parallelism raises decode speed
The benchmark compares two ways of dividing work. Pipeline parallelism assigns different layers to different Macs, so one system passes its output to the next. Tensor parallelism splits computation within layers across the machines, enabling them to work on each token together but requiring frequent data exchanges.
For Muse Spark Open, Apple reports 15.74 tokens per second with two Macs using pipeline parallelism, 24.99 with two using tensor parallelism, and 36.73 with four using tensor parallelism. The four-Mac result is more than twice the two-Mac pipeline figure, but it is not a comparison against a single-Mac run: Apple identifies the two-Mac pipeline setup as the baseline for this model in its speedup analysis.
The same table gives single-machine results for two smaller models. At the same 8,192-token context, four M5 Ultra systems generate Mistral Medium 3.5 at 23.37 tokens per second, compared with 7.70 on one system. Qwen 3.8 reaches 50.89 tokens per second on four, against 19.69 on one.
Thunderbolt RDMA carries the exchanges
Apple attributes the cluster’s low-latency communication to RDMA over Thunderbolt 5. MLX uses the JACCL communication backend for collective operations between machines. The benchmark setup connects Mac Studios with two Thunderbolt cables between each pair to increase link bandwidth.
That connection matters especially for tensor parallelism, which requires devices to exchange information repeatedly as they process each token. Apple says the overhead can offset some gains: at very long contexts, two-machine tensor parallelism for Muse Spark Open trails two-machine pipeline parallelism during prompt processing. With four M5 Ultra systems, the company says, tensor parallelism pays off for decoding even at the larger context length.
MLX can serve the model to local tools
Apple’s post also shows how to run Muse Spark Open through MLX’s server interface, then connect it to OpenCode as a local model provider. That gives developers a path from a benchmark to an agentic coding setup, with inference running on their Mac cluster rather than through a remote model endpoint.
The release makes a large model runnable on a desk-scale cluster, but the reported rate is a specific vendor measurement under a defined setup—not a general promise of response speed across tasks, contexts or software. Apple’s clearest result is narrower: four M5 Ultra Mac Studios, using tensor parallelism over Thunderbolt RDMA, decode Muse Spark Open at 36.73 tokens per second at an 8,192-token context.