Muse Spark Open measured: 25.2 tok/s on 3x A100

Published 2026-10-16

We ran Meta's Muse Spark Open (816B total / 42B active) on three A100 80GB cards with Unsloth's 175.6 GB 1-bit file: 25.2 tok/s chat-short decode, ~168 GiB peak VRAM. That roughly matches Apple's unverified ~25 tok/s claim for two Mac Studios. Full page and FAQ: /m/muse-spark-open.

This is Unsloth's smallest 1-bit quant (UD-IQ1_S) on llama.cpp 2851b3bc (build 11675) at 16K context, whole model on GPU, run cost $1.52. Every figure is a single run, not a median. Time to first token (11.2 s) and prefill (288.7 tok/s) come from suite-v1 chat-long: this model's tokenizer counted 3,227 prompt tokens (suite target 4,096). We measure speed and memory, not output quality.

Measured workloads

Unsloth UD-IQ1_S 1-bit (175.6 GB), 3x A100 80GB, llama.cpp 2851b3bc, 16K context. One run per workload, not a median. Chat-long prefill 288.7 tok/s and TTFT 11.2 s on a 3,227-token prompt.
Workload	Decode	Run
`chat-short`	25.2 tok/s	r_fsd9qjmmna0
`chat-long`	25.5 tok/s	r_fsd9qjmmna0
`concurrent-decode`	26.3 tok/s	r_fsd9qjmmna0
`agent-trace`	22.8 tok/s	r_fsd9qjmmna0

What about Mac Studios?

Apple/MLX claim about 25 tok/s on two M5 Ultra Mac Studios and about 36 on four. We haven't verified those and keep them out of our tables; the 25.2 tok/s match is speed only, on different hardware and a different quant. If you have a multi-Studio setup, submit an MLX run with the request button and we'll set it against the claims.