Breaking01 / AI
Four Mac Studios run 816B Muse Spark Open with 4-bit weights
The figures are Apple's own measurements, and the company did not report how 4-bit quantisation affected answer quality.
Apple reported on 15 Oct 2026 that four Mac Studio computers ran Muse Spark Open at 36.73 tokens per second (8,192-token context) in its own benchmark tests, published on its research blog; no independent replication has been published.
The measurements used the mlx_lm.benchmark tool on Mac Studio machines with the M5 Ultra chip, according to Apple's research blog. Model sizes were given as 816 billion parameters for Muse Spark Open (42 billion active), 128 billion for Mistral Medium 3.5 and 27 billion for Qwen 3.8.
Communication between the machines ran over an RDMA (remote direct memory access) driver based on Thunderbolt 5, Apple said, and tensor parallelisation split each layer of the model across the devices. Time to first token on the four-machine setup was 40.6 seconds with two cables per connection and about 50.5 seconds with one (65,536-token context).
Apple put the two-machine speed for the same model at 24.99 tokens per second (tensor parallelisation, 8,192-token context). On a single machine, generation was 1.6 times faster on the M5 Ultra than on the M3 Ultra (Qwen 3.8 and Mistral Medium 3.5). Time to first token was up to 4.6 times faster than on the M3 Ultra (same two models, 8,192- and 65,536-token contexts).
Muse Spark Open occupies about 493GB even with 4-bit quantisation (NVFP4), a technique that stores model values at lower precision to reduce memory use, Apple said. Once memory for intermediate computation is added, the model does not fit on a single 512GB machine. At a 65,536-token context, two machines using tensor parallelisation lagged two using pipeline parallelisation, which gives each machine a consecutive block of layers, during prefill, the stage that processes the input context.
Apple's earlier post covered the M5 architecture and the benefits of its GPU Neural Accelerators, while the new one reports figures for runs split across several machines. No replication outside Apple has been published, and Apple's post does not report how 4-bit quantisation affected the quality of Muse Spark Open's answers.