Apple M3 Max (64GB) Llama-3.3-70B Q4_K_M decode
8.2 tok/s decode at 258 ctx, falling to 6.1 at 32k ctx
Fields
id
benchmarks/comm-apple-m3-max-llama33-70b-4bit-decodeMetric
decode_tok_s
Value
8.2
Unit
tok/s decode at 258 ctx, falling to 6.1 at 32k ctx
Model
Llama-3.3-70B (dense) Q4_K_M
Methodology
Silicon Score Llama-3.3-70B machine row; context-depth degrades decode on ~300-400 GB/s part
Measured by
third_party
Thermal state
unknown
notes
Context length crushes decode on the Max tier: 8.2 t/s @258 ctx -> 6.1 t/s @32k ctx. Bandwidth 300-400 GB/s depending on config.
Notes
Context length crushes decode on the Max tier: 8.2 t/s @258 ctx -> 6.1 t/s @32k ctx. Bandwidth 300-400 GB/s depending on config.