Qwen3.5-4B · Q3_K_L
This page aggregates 2 real-world runs of Qwen3.5-4B (Q3_K_L) on Apple M5 Max with Strata, contributed by 1 independent source platforms; metrics are averages of published measurements.
354.31 tok/s
Decode
Decode speed
1,310.91 tok/s
Prefill
Prefill speed
1.03 s
TTFT
Time to first token
8.15 GB
VRAM
VRAM usage
2
Measured runs
1
Independent sources
Other
Source platforms
2 days ago
Last verified
Performance
- llama.cpp · IQ3_S Decode 279.98 · Prefill 1354 ·
- Strata · Q3_K_L (current)Decode 354.31 · Prefill 1310.91 ·
- Strata · Q5_K_S Decode 279.1 · Prefill 1192.93 ·
- vLLM · IQ2_M Decode 364.07 · Prefill 1250.03 ·
- vLLM · Q4_K_S Decode 256.46 · Prefill 848.78 ·
- Strata · Q5_K_M Decode 272.64 · Prefill 963.9 ·
Core figures
- Decode (avg)
- 354.31 tok/s
- Prefill (avg)
- 1,310.91 tok/s
- TTFT (avg)
- 1.03 s
- VRAM (avg)
- 8.15 GB
- MTP acceptance rate
- 0.62
- TTFB
- 2.84 GB
- Power draw
- 133.6 W
Configuration
Member-level fields are taken from the most recent run
- Model
- Qwen3.5-4B
- Quantization
- Q3_K_L
- Framework
- Strata
- Version
- v0.25
- Context length
- 4096 tokens
- Batch size
- 256
- GPU layers
- 12
- Flash Attention
- Off
Hardware
Nominal and measured figures are shown side by side; whether it runs is the reader's call
- GPU
- Apple M5 Max
- Nominal VRAM
- 128 GB
- Measured VRAM (avg)
- 8.15 GB
- OS
- Ubuntu 22.04
- Driver
- 570.86
- CUDA
- 12.8
- Power draw
- 161.9 W
Sources & evidence
2 measured records in total, each traceable to its original source
v0.25 · Ubuntu 22.04 · CUDA 12.8 · 4096 ctx
354.71 tok/s
Decode
1,193.74 tok/s
Prefill
1.68 s
TTFT
8.15 GB
VRAM
0.54
MTP
161.9 W
Power draw
L2 batch seed 2026-10-16: Strata on hw_id=28 running model_id=48 at Q3_K_L
v0.31 · macOS 15.0 · CUDA N/A · 2048 ctx
353.9 tok/s
Decode
1,428.08 tok/s
Prefill
0.39 s
TTFT
8.15 GB
VRAM
0.7
MTP
105.3 W
Power draw
L2 batch seed 2026-10-13: Strata on hw_id=28 running model_id=48 at Q3_K_L