← Back to community benchmarks

Qwen3___5-4B

M3 Ultra (80c) · 512 GB · 4bit · 2026-03-09
Performance
8k
tokens
2,542
PP tok/s
133.1
TG tok/s
3223
TTFT (ms)
4.6
Peak mem (GB)
Hardware
Chip M3 Ultra (80c)
Memory 512 GB
GPU Cores 80
Software
oMLX v0.2.6
macOS macOS 26.3.1
Context 8,192
Performance by Context Length
Context PP tok/s TG tok/s Peak Mem
1k 2,581 152.6 4.0 GB view
4k 2,608 143.4 4.2 GB view
8k 2,542 133.1 4.6 GB current
16k 2,385 116.4 5.3 GB view
32k 2,116 90.5 6.9 GB view
64k 1,701 61.3 8.6 GB view
Batching Results
Batch Size TG tok/s Speedup
152.6 1.00×
263.9 1.73×
371.8 2.44×