Apple: four Mac Studios run Muse Spark Open at 36 tokens a second

Apple's machine learning research team published a post on 15 October 2026 showing how to run Meta's Muse Spark Open, an 816-billion-parameter open-weight model with 42 billion parameters active per token, across several Mac Studio computers. Apple reports that four Mac Studios with M5 Ultra chips generate 36.73 tokens per second at an 8,192-token context. The post is Apple's own benchmark of its own software, MLX, released the same day.

What Apple measured

Apple tested three models: Muse Spark Open (816B total, 42B active, evaluated in 4-bit NVFP4), Mistral Medium 3.5 (128B, FP8) and Qwen 3.8 (27B, BF16). It ran them with the mlx_lm.benchmark tool on Mac Studios with up to 512GB of unified memory, comparing the M5 Ultra against the previous M3 Ultra generation.

• Single machine: Apple reports the M5 Ultra decodes 1.6x faster than the M3 Ultra (1.2 TB/s of memory bandwidth against 819 GB/s) and produces the first token up to 4.6x faster. Peak power was about 300W for the M5 Ultra and 200W for the M3 Ultra.
• Muse Spark Open decode at 8,192 tokens: 15.74 tokens/s on two machines splitting the model by layers (pipeline), 24.99 on two splitting each layer (tensor), and 36.73 on four in tensor mode.
• Mistral Medium 3.5 decode on the M5 Ultra: 7.70 tokens/s on one machine and 23.37 on four. Qwen 3.8: 19.69 on one and 50.89 on four.

Why it takes a cluster

Apple says Muse Spark Open would need more than 1.6TB of memory for parameters alone in its native BF16 format. At 4-bit that falls to about 493GB, and with working memory added it still does not fit on one machine. The Macs are linked with Thunderbolt 5 cables using an RDMA driver introduced in macOS 26.2; Apple says two cables per link cut Muse Spark Open's time to first token on four machines at a 65,536-token context from about 50.5 seconds to 40.6 seconds.

What the post leaves out

Apple gives no price for the cluster, no figures for several users at once, and no quality results for the 4-bit model. Its comparison set is uneven: Muse Spark Open is run at 4-bit while Qwen 3.8 and Mistral Medium 3.5 run at native precision. In the Hacker News thread, commenter peri-cl argued that this sandbags Qwen and that Muse Spark Open would not make their top ten models for a 512GB machine; another, ks2048, observed that the arrangement sends hardware spend to Apple for models Meta paid to build.

What a team can take from it

For a company that cannot send data to a hosted model, the useful facts are the memory arithmetic and the scaling. Apple reports that decode speed rises with each added machine when tensor parallelism and fast links are used, and that dense models scaled better than this mixture-of-experts one. That is a single-user speed figure. Before budgeting four workstations, measure tokens per second with your own prompts and context lengths, with several users at once, and compare answer quality at 4-bit against the model you run today.