Back to search results
Qwen3-4B-Thinking-2507 · Q2_K
NVIDIA RTX 509032GBllama.cpp v0.33L2 Cross-framework verified
This page aggregates 1 real-world runs of Qwen3-4B-Thinking-2507 (Q2_K) on NVIDIA RTX 5090 with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.
392.48 tok/s
Decode
Decode speed
1,604.11 tok/s
Prefill
Prefill speed
0.17 s
TTFT
Time to first token
6.72 GB
VRAM
VRAM usage
L2 Cross-framework verified
1
Measured runs
1
Independent sources
Other
Source platforms
2 days ago
Last verified
Performance
- llama.cpp · Q2_K (current)Decode 392.48 · Prefill 1604.11 ·
- llama.cpp · Q8_0 Decode 184.55 · Prefill 803.14 ·
- vLLM · IQ2_M Decode 386.87 · Prefill 1842.03 ·
- vLLM · Q3_K_L Decode 337.16 · Prefill 1247.18 ·
- vLLM · Q5_K_M Decode 507.7 · Prefill 3807.8 ·
- vLLM · Q4_K_M Decode 634.7 · Prefill 4760.3 ·
- llama.cpp · Q5_K_M Decode 465.9 · Prefill 3494.3 ·
- llama.cpp · Q4_K_M Decode 582.4 · Prefill 4368 ·
Core figures
- Decode (avg)
- 392.48 tok/s
- Prefill (avg)
- 1,604.11 tok/s
- TTFT (avg)
- 0.17 s
- VRAM (avg)
- 6.72 GB
- MTP acceptance rate
- —
- TTFB
- 3.14 GB
- Power draw
- 308.7 W
Configuration
Member-level fields are taken from the most recent run
- Model
- Qwen3-4B-Thinking-2507
- Quantization
- Q2_K
- Framework
- llama.cpp
- Version
- v0.33
- Context length
- 2048 tokens
- Batch size
- 256
- 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
- NVIDIA RTX 5090
- Nominal VRAM
- 32 GB
- Measured VRAM (avg)
- 6.72 GB
- OS
- Ubuntu 24.04
- Driver
- 565.57
- CUDA
- 12.8
- Power draw
- 308.7 W
Sources & evidence
1 measured records in total, each traceable to its original source
v0.33 · Ubuntu 24.04 · CUDA 12.8 · 2048 ctx
392.48 tok/s
Decode
1,604.11 tok/s
Prefill
0.17 s
TTFT
6.72 GB
VRAM
—
MTP
308.7 W
Power draw
L2 batch seed 2026-10-16: llama.cpp on hw_id=13 running model_id=57 at Q2_K