Ollama (llama.cpp backend) on MacBook Pro M3 Max 128GB: Qwen3.8-27B Q4_K_M decode tokens/s
12.0 tok/s (batch 1, short prompt decode)
Fields
id
benchmarks/ollama-m3-max-qwen3-8-27b-llamacpp-decode-tok-sMetric
decode_tok_s
Value
12
Unit
tok/s (batch 1, short prompt decode)
Model
Qwen3.8 27B
Methodology
MacBook Pro M3 Max (16-core CPU / 40-core GPU, 128 GB unified memory at 400 GB/s), macOS 26.6.2. Ollama running with llama.cpp engine backend, Q4_K_M GGUF format, batch size 1, temp 0, reasoning off, repeated 4x with variance under 2%.
Measured by
third_party
Reproducible
yes
notes
Direct head-to-head empirical measurement on M3 Max 128GB (400 GB/s). Under llama.cpp backend, Qwen3.8-27B achieves 12.0 tok/s decode, trailing MLX (42.5 tok/s) by 3.54x on the identical hardware due to inefficient handling of Gated DeltaNet (GDN) hybrid layers in GGML Metal kernels.
Notes
Direct head-to-head empirical measurement on M3 Max 128GB (400 GB/s). Under llama.cpp backend, Qwen3.8-27B achieves 12.0 tok/s decode, trailing MLX (42.5 tok/s) by 3.54x on the identical hardware due to inefficient handling of Gated DeltaNet (GDN) hybrid layers in GGML Metal kernels.