Apple M1 Ultra (64GB), llama.cpp/LM Studio: Llama-3.3-70B 4-bit decode tok/s (Silicon Score measured)
12.6 decode tok/s (single-stream, 4-bit, 4K ctx, LM Studio)
Fields
id
benchmarks/comm-apple-m1-ultra-llama33-70b-4bit-decodeMetric
decode_tok_s
Value
12.6
Unit
decode tok/s (single-stream, 4-bit, 4K ctx, LM Studio)
Model
Llama-3.3-70B (dense) 4-bit
Methodology
THIRD-PARTY MEASURED (Silicon Score Llama 3.3 70B model page + chip rows; llama.cpp discussion #4167 maintainer benchmark table). M1 Ultra 64GB, ~800 GB/s unified memory: 12.6 tok/s on 70B 4-bit at 4K ctx. On the canonical M-series table (discussion #4167) M1 Ultra does 83.7 tok/s Q4_0 on a 7B-class model - decode scales almost perfectly with memory bandwidth. Cheapest ticket to a usable 70B: used Mac Studio M1 Ultra 128GB ~$2,200-3,000.
Measured by
third_party
Reproducible
yes
notes
The best $/tok for big dense models among the 'lots of memory for not-crazy money' paths. The whole 128GB capacity is usable by the GPU - no VRAM ceiling. Idles ~15-20 W, ~90-150 W under sustained inference, whisper-quiet. Used pricing is DRAM-shortage-inflated (+10-15% since early 2026). Sibling rows measured: M2 Max 38c 64GB 8.8 t/s; M3 Max 64GB 8.2->6.1 Q4_K_M (258->32k ctx); M4 Max 40c 128GB 11.8 (4bit) / 6.5 (8bit); M3 Ultra 512GB 15.5 (4bit) / 8.5 (8bit).
Notes
The best $/tok for big dense models among the 'lots of memory for not-crazy money' paths. The whole 128GB capacity is usable by the GPU - no VRAM ceiling. Idles ~15-20 W, ~90-150 W under sustained inference, whisper-quiet. Used pricing is DRAM-shortage-inflated (+10-15% since early 2026). Sibling rows measured: M2 Max 38c 64GB 8.8 t/s; M3 Max 64GB 8.2->6.1 Q4_K_M (258->32k ctx); M4 Max 40c 128GB 11.8 (4bit) / 6.5 (8bit); M3 Ultra 512GB 15.5 (4bit) / 8.5 (8bit).