llama.cpp Metal on Apple M6 (Mac mini, 12-core GPU, 32 GB): llama 7B F16 (12.55 GiB, 6.74B) pp512 (discussion #4167)
952.53 tok/s (pp512, per-request, batch 1)
Fields
id
benchmarks/llamacpp-metal-m6-f16-pp512Metric
prefill_tok_s
Value
952.53
Unit
tok/s (pp512, per-request, batch 1)
Model
llama 7B F16 (12.55 GiB, 6.74B)
Methodology
COMMUNITY llama-bench (llama.cpp discussion #4167, comment by @jhen0409 dated 2026-10-01). Mac mini M6, 12-core GPU, 32 GB unified memory, reporter-stated bandwidth 170 GB/s, macOS 27.0.1. Backend BLAS,MTL, 2 threads. Build commit 0c1e570. Mean of the default repeats. Per-request batch-1, NOT aggregate. Model: llama 7B F16 (12.55 GiB, 6.74B).
Measured by
third_party
Reproducible
yes
notes
First measured coverage for the base M6. The pinned scoreboard table lists the same figures under commit c1d0e7a; the reporter's own paste cites build 0c1e570, which this record uses. format_id left null: F16 GGUF has no quantization record (same convention as the other llama-2-7b F16 rows).
Notes
First measured coverage for the base M6. The pinned scoreboard table lists the same figures under commit c1d0e7a; the reporter's own paste cites build 0c1e570, which this record uses. format_id left null: F16 GGUF has no quantization record (same convention as the other llama-2-7b F16 rows).