llama.cpp Metal on Apple M2 Max (30-core GPU, 32 GB): Qwen3.6-35B-A3B Q4_K_M pp4096
671.4 tok/s (pp4096, per-request, batch 1)
Fields
id
benchmarks/llamacpp-metal-m2-max-30core-qwen36-35b-a3b-q4km-pp4096Metric
prefill_tok_s
Value
671.4
Unit
tok/s (pp4096, per-request, batch 1)
Model
Qwen3.6-35B-A3B Q4_K_M (20.22 GiB, MoE, 3B active)
Methodology
COMMUNITY llama-bench, llama.cpp discussion #4167 comment 18555918. Posted by @dmaynard51, comment 18555918 (2026-09-22): M2 Max, 30-core GPU, 32 GB, GPU otherwise idle. This is NOT the thread's standard llama-7b pp512/tg128 cell: prompt length is 4096 and the models are Qwen3.6-35B-A3B Q4_K_M and Qwen3.8-27B IQ3_S. llama-bench, single stream, batch 1. Build not printed in the comment. The MTP rows in the same comment are ollama runs with a draft head and are not recorded here. Reported 671.40 +/- 11.20. Per-request, NOT aggregate.
Measured by
third_party
Reproducible
yes
notes
20.22 GiB file on a 32 GB machine, so context headroom is small. The MoE prefill is 6.3x the dense 27B row below despite the larger file.
Notes
20.22 GiB file on a 32 GB machine, so context headroom is small. The MoE prefill is 6.3x the dense 27B row below despite the larger file.