Qwen3-32B · Q4_K_M
This page aggregates 3 real-world runs of Qwen3-32B (Q4_K_M) on NVIDIA RTX 4090 with llama.cpp, contributed by 2 independent source platforms; metrics are averages of published measurements.
18.4 tok/s
Decode
Decode speed
216.93 tok/s
Prefill
Prefill speed
1.19 s
TTFT
Time to first token
19.5 GB
VRAM
VRAM usage
3
Measured runs
2
Independent sources
GitHub, Reddit
Source platforms
19 days ago
Last verified
Performance
- llama.cpp · Q4_K_M (current)Decode 18.4 · Prefill 216.93 ·
- Ollama · Q4_K_M Decode 17.4 · Prefill 198.2 ·
Core figures
- Decode (avg)
- 18.4 tok/s
- Prefill (avg)
- 216.93 tok/s
- TTFT (avg)
- 1.19 s
- VRAM (avg)
- 19.5 GB
- MTP acceptance rate
- —
- TTFB
- — GB
- Power draw
- 347.17 W
Configuration
Member-level fields are taken from the most recent run
- Model
- Qwen3-32B
- Quantization
- Q4_K_M
- Framework
- llama.cpp
- Version
- b4567
- Context length
- 8192 tokens
- Batch size
- 512
- GPU layers
- 99
- Flash Attention
- On
Hardware
Nominal and measured figures are shown side by side; whether it runs is the reader's call
- GPU
- NVIDIA RTX 4090
- Nominal VRAM
- 24 GB
- Measured VRAM (avg)
- 19.5 GB
- OS
- Ubuntu 24.04
- Driver
- 550.107.02
- CUDA
- 12.4
- Power draw
- 342.5 W
Sources & evidence
3 measured records in total, each traceable to its original source
b4567 · Ubuntu 24.04 · CUDA 12.4 · 8192 ctx
18 tok/s
Decode
210.5 tok/s
Prefill
1.24 s
TTFT
19.5 GB
VRAM
—
MTP
342.5 W
Power draw
Q4_K_M, 8192 ctx, single 4090: ~18 tok/s decode.
b4602 · Ubuntu 24.04 · CUDA 12.4 · 8192 ctx
18.6 tok/s
Decode
218 tok/s
Prefill
1.18 s
TTFT
19.4 GB
VRAM
—
MTP
348 W
Power draw
复现成功，decode 18.6 tok/s，与本机一致。
b4602 · Windows 11 24H2 · CUDA 12.4 · 16384 ctx
18.6 tok/s
Decode
222.3 tok/s
Prefill
1.15 s
TTFT
19.6 GB
VRAM
—
MTP
351 W
Power draw
llama.cpp 与 Ollama 两条路径均复现，18-19 tok/s。