SiliconScore community row: Apple M3 Ultra, MLX, Qwen3.5-397B-A17B q4.1bit MLX, decode 40.2 t/s
40.2 tok/s avg decode (single-request)
Fields
id
benchmarks/ss-m3-ultra-qwen35-397b-a17b-41bit-decodeMetric
decode_tok_s
Value
40.2
Unit
tok/s avg decode (single-request)
Model
Qwen3.5-397B-A17B (q4.1bit MLX)
Methodology
COMMUNITY-UNVERIFIED row aggregated by siliconscore.com M3 Ultra chip page (accessed 2026-10-15); underlying evidence is the linked r/LocalLLaMA/HF measurement. MLX runtime, 4-bit-ish quant as named. avg decode tok/s single-request. Low-confidence forum-class evidence. RAM tier: 512 GB. A ~400B-class MoE at 40 tok/s is the M3 Ultra's headline MoE row; source is the inferencerlabs HF model card.
Measured by
third_party
Reproducible
no