← Back to community benchmarks

Qwen3.5-27B-heretic

M3 Ultra (80c) · 512 GB · 8bit · 2026-03-09
Performance
64k
tokens
331.4
PP tok/s
15.0
TG tok/s
197764
TTFT (ms)
36.2
Peak mem (GB)
Hardware
Chip M3 Ultra (80c)
Memory 512 GB
GPU Cores 80
Software
oMLX v0.2.6
macOS macOS 26.3
Context 65,536
Performance by Context Length
Context PP tok/s TG tok/s Peak Mem
1k 413.4 21.6 29.1 GB view
4k 421.8 21.0 29.5 GB view
8k 416.8 20.4 29.9 GB view
16k 403.9 19.2 30.8 GB view
32k 378.7 17.5 32.6 GB view
64k 331.4 15.0 36.2 GB current
Batching Results
Batch Size TG tok/s Speedup
21.6 1.00×
38.4 1.78×
57.3 2.65×
70.7 3.27×