verifiedmedium · 0.80updated 2026-10-15

LLMCheck: Apple M5 Ultra (96GB), MLX, Llama 3.3 70B Q4_K_M decode

24.0 tok/s (decode, per-request)

Fields

benchmarks/llmcheck-m5-ultra-llama33-70b-q4km-mlx-decode

Metric

decode_tok_s

Value

Unit

tok/s (decode, per-request)

Model

Llama 3.3 70B

Methodology

THIRD-PARTY MEASURED (LLMCheck open dataset, dated 2026-08). M5 Ultra 96GB, MLX, Llama 3.3 70B Q4_K_M, per-request decode.

Measured by

third_party

Reproducible

yes

notes

Roughly 2x the M5 Max 70B figure (24 vs 12 tok/s) — tracks the 1.2 TB/s vs 614 GB/s bandwidth ratio almost exactly, textbook bandwidth-bound decode.

Roughly 2x the M5 Max 70B figure (24 vs 12 tok/s) — tracks the 1.2 TB/s vs 614 GB/s bandwidth ratio almost exactly, textbook bandwidth-bound decode.