llama.cpp Metal on Apple M2 Max (30-core GPU, 32 GB): Qwen3.6-35B-A3B Q4_K_M tg128
58.56 tok/s (tg128, per-request, batch 1)
Fields
id
benchmarks/llamacpp-metal-m2-max-30core-qwen36-35b-a3b-q4km-tg128Metric
decode_tok_s
Value
58.56
Unit
tok/s (tg128, 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 58.56 +/- 0.82. Per-request, NOT aggregate.
Measured by
third_party
Reproducible
yes
notes
5.4x the dense 27B decode on the same chip and the same comment.
Notes
5.4x the dense 27B decode on the same chip and the same comment.