llama.cpp Metal on Apple M2 Max (30-core GPU, 32 GB): Qwen3.8-27B IQ3_S tg128
10.82 tok/s (tg128, per-request, batch 1)
Fields
id
benchmarks/llamacpp-metal-m2-max-30core-qwen38-27b-iq3s-tg128Metric
decode_tok_s
Value
10.82
Unit
tok/s (tg128, per-request, batch 1)
Model
Qwen3.8-27B (dense) IQ3_S (11.20 GiB)
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 10.82 +/- 0.06. Per-request, NOT aggregate.
Measured by
third_party
Reproducible
yes
notes
Dense decode on the 30-core M2 Max. The same comment's MTP sweep (draft-mtp n=2 at 34.6 vs 46.2 without) is a different runtime and is not this row.
Notes
Dense decode on the 30-core M2 Max. The same comment's MTP sweep (draft-mtp n=2 at 34.6 vs 46.2 without) is a different runtime and is not this row.