Qwen3-4B · IQ2_XS
This page aggregates 2 real-world runs of Qwen3-4B (IQ2_XS) on Apple M5 Max with Ollama, contributed by 1 independent source platforms; metrics are averages of published measurements.
308.45 tok/s
Decode
Decode speed
1,307.27 tok/s
Prefill
Prefill speed
1.61 s
TTFT
Time to first token
6.13 GB
VRAM
VRAM usage
2
Measured runs
1
Independent sources
Other
Source platforms
2 days ago
Last verified
Performance
- llama.cpp · Q3_K_L Decode 253.55 · Prefill 980.61 ·
- Ollama · IQ2_XS (current)Decode 308.45 · Prefill 1307.27 ·
- Ollama · IQ3_XS Decode 249.13 · Prefill 1132.99 ·
- Ollama · Q5_K_M Decode 194.79 · Prefill 989.8 ·
- Strata · IQ4_XS Decode 357.04 · Prefill 1403.97 ·
Core figures
- Decode (avg)
- 308.45 tok/s
- Prefill (avg)
- 1,307.27 tok/s
- TTFT (avg)
- 1.61 s
- VRAM (avg)
- 6.13 GB
- MTP acceptance rate
- —
- TTFB
- 2.47 GB
- Power draw
- 115.8 W
Configuration
Member-level fields are taken from the most recent run
- Model
- Qwen3-4B
- Quantization
- IQ2_XS
- Framework
- Ollama
- Version
- v0.21
- Context length
- 4096 tokens
- Batch size
- 512
- GPU layers
- 99
- 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)
- 6.13 GB
- OS
- Ubuntu 22.04
- Driver
- 565.57
- CUDA
- 12.8
- Power draw
- 159.7 W
Sources & evidence
2 measured records in total, each traceable to its original source
v0.21 · Ubuntu 22.04 · CUDA 12.8 · 4096 ctx
287.42 tok/s
Decode
1,006.86 tok/s
Prefill
2.34 s
TTFT
6.13 GB
VRAM
—
MTP
159.7 W
Power draw
L2 batch seed 2026-10-16: Ollama on hw_id=28 running model_id=54 at IQ2_XS
v0.34 · macOS 15.0 · CUDA N/A · 8192 ctx
329.47 tok/s
Decode
1,607.68 tok/s
Prefill
0.88 s
TTFT
6.13 GB
VRAM
—
MTP
71.9 W
Power draw
L2 batch seed 2026-10-13: Ollama on hw_id=28 running model_id=54 at IQ2_XS