Qwen3-4B-Thinking-2507 · Q5_K_M
This page aggregates 2 real-world runs of Qwen3-4B-Thinking-2507 (Q5_K_M) on AMD Ryzen AI Max+ 395 with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.
107.36 tok/s
Decode
Decode speed
478 tok/s
Prefill
Prefill speed
0.7 s
TTFT
Time to first token
9.91 GB
VRAM
VRAM usage
2
Measured runs
1
Independent sources
Other
Source platforms
2 days ago
Last verified
Performance
- llama.cpp · Q5_K_M (current)Decode 107.36 · Prefill 478 ·
- llama.cpp · Q8_0 Decode 78.29 · Prefill 263.52 ·
- Ollama · Q3_K_S Decode 145.01 · Prefill 581.64 ·
- llama.cpp · IQ2_M Decode 178.19 · Prefill 583.74 ·
- Strata · Q3_K_L Decode 190.53 · Prefill 851.53 ·
Core figures
- Decode (avg)
- 107.36 tok/s
- Prefill (avg)
- 478 tok/s
- TTFT (avg)
- 0.7 s
- VRAM (avg)
- 9.91 GB
- MTP acceptance rate
- —
- TTFB
- 0.86 GB
- Power draw
- 86.15 W
Configuration
Member-level fields are taken from the most recent run
- Model
- Qwen3-4B-Thinking-2507
- Quantization
- Q5_K_M
- Framework
- llama.cpp
- Version
- v0.47
- Context length
- 16384 tokens
- Batch size
- 512
- GPU layers
- 20
- Flash Attention
- On
Hardware
Nominal and measured figures are shown side by side; whether it runs is the reader's call
- GPU
- AMD Ryzen AI Max+ 395
- Nominal VRAM
- 128 GB
- Measured VRAM (avg)
- 9.91 GB
- OS
- Ubuntu 22.04
- Driver
- 565.57
- CUDA
- N/A
- Power draw
- 111.3 W
Sources & evidence
2 measured records in total, each traceable to its original source
v0.47 · Ubuntu 22.04 · CUDA N/A · 16384 ctx
105.16 tok/s
Decode
413.34 tok/s
Prefill
0.62 s
TTFT
9.91 GB
VRAM
—
MTP
111.3 W
Power draw
L2 batch seed 2026-10-16: llama.cpp on hw_id=24 running model_id=57 at Q5_K_M
v0.45 · Windows 11 23H2 · CUDA N/A · 2048 ctx
109.56 tok/s
Decode
542.66 tok/s
Prefill
0.77 s
TTFT
9.91 GB
VRAM
—
MTP
61 W
Power draw
L2 batch seed 2026-10-13: llama.cpp on hw_id=24 running model_id=57 at Q5_K_M