Apple M3 Max (64GB), llama.cpp: Llama-3.3-70B Q4_K_M decode tok/s, context-scaled (Silicon Score measured)
6.1 decode tok/s (single-stream, Q4_K_M, 32K ctx)
Fields
id
benchmarks/comm-apple-m3-max-llama33-70b-q4km-decodeMetric
decode_tok_s
Value
6.1
Unit
decode tok/s (single-stream, Q4_K_M, 32K ctx)
Model
Llama-3.3-70B (dense) Q4_K_M
Methodology
THIRD-PARTY MEASURED (Silicon Score Llama 3.3 70B model page + chip rows). M3 Max 64GB, ~400 GB/s (16-core CPU / 40-core GPU configuration): 8.2 tok/s at 258-token context falling to 6.1 tok/s at 32K context on Llama-3.3-70B Q4_K_M. Recorded value is the 32K-ctx floor; the 258-ctx figure (8.2) is the short-context ceiling - context length crushes decode speed as KV reads grow to rival weight reads.
Measured by
third_party
Reproducible
yes
notes
CREATED 2026-10-15 Phase-A residual gap-fill - the 8.2->6.1 ctx-scaled pair was stranded in comm-apple-m1-ultra-llama33-70b-4bit-decode's notes with no record of its own. Value field carries the LONG-context (32K) number as the conservative planning figure; methodology carries the short-context 8.2. This is the clearest single illustration in the Apple block that 'a tok/s number' is meaningless without its context length: the same part + model + quant swings 35% from 258 to 32K tokens.
Notes
CREATED 2026-10-15 Phase-A residual gap-fill - the 8.2->6.1 ctx-scaled pair was stranded in comm-apple-m1-ultra-llama33-70b-4bit-decode's notes with no record of its own. Value field carries the LONG-context (32K) number as the conservative planning figure; methodology carries the short-context 8.2. This is the clearest single illustration in the Apple block that 'a tok/s number' is meaningless without its context length: the same part + model + quant swings 35% from 258 to 32K tokens.