Muse Spark Open 816B (42B active)
4× M5 Ultra · MLX · 8,192 ctx
- generation:
- 36.7 tokens/s
- quant:
- NVFP4 (NVFP4)
Reported by the source; GPU count, offloading and concurrent requests can change these figures. Prompt-processing speed is shown only when the source reports it; generation speed cannot tell us time to first token. Check the full setup before comparing.
User reports 36.73 t/s generation. Source evidence (verbatim): Source: https://machinelearning.apple.com/research/mlx-distributed-inference # Apple vendor Muse Spark Open NVFP4 distributed Mac Studio benchmark Apple first-party research, October 15, 2026. Muse Spark Open: 816B total,42B active parameters, NVFP4. Four M5 Ultra Mac Studios, each 512GB unified memory; tensor parallelism, jaccl RDMA over Thunderbolt5, two cables per pair in mesh. Vendor native decode 36.73 tokens/s, TTFT 4.8 seconds in initially rendered table; chart specifies 8192 context. MLX / mlx_lm.benchmark: 8192 input tokens,128 generated tokens,batch one (-p 8192 -g 128 -b 1); MLX_METAL_FAST_SYNCH=1. Do not convert TTFT into prefill throughput. Tables also toggle 65536; that state not independently extracted. Exact CPU/GPU core counts, tested runtime version/commit, artifact revision, KV precision and sampling unknown. NVFP4 model footprint about493GB is not cluster resident/peak memory. Tested repository named by benchmark command and publicly verified: https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B-NVFP4 . Revision unknown; no download/license acceptance performed. Two-node pipeline/tensor results are separate configurations. Full source and caveats must remain public.