Back to search results

Qwen3-8B · Q4_K_M

Start a discussion
NVIDIA RTX 409024GBllama.cpp v0.48L2 Cross-framework verified

This page aggregates 2 real-world runs of Qwen3-8B (Q4_K_M) on NVIDIA RTX 4090 with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.

142.71 tok/s

Decode

Decode speed

813.73 tok/s

Prefill

Prefill speed

1.2 s

TTFT

Time to first token

10.64 GB

VRAM

VRAM usage

L2 Cross-framework verified

2

Measured runs

1

Independent sources

Other

Source platforms

2026-10-19

Last verified

Performance

  1. llama.cpp · Q4_K_M (current)Decode 142.71 · Prefill 813.73 ·
  2. vLLM · IQ3_S Decode 161.14 · Prefill 813.02 ·
  3. llama.cpp · IQ1_S Decode 187.88 · Prefill 983.15 ·
  4. llama.cpp · Q8_0 Decode 87.18 · Prefill 465.1 ·
  5. vLLM · Q2_K Decode 164.91 · Prefill 671.93 ·
  6. llama.cpp · Q6_K Decode 88.33 · Prefill 428.17 ·
  7. vLLM · Q4_K_S Decode 126.59 · Prefill 467.91 ·
  8. vLLM · Q8_0 Decode 76.65 · Prefill 513.21 ·
  9. Ollama · Q6_K Decode 82.66 · Prefill 332.47 ·
  10. llama.cpp · IQ4_XS Decode 112 · Prefill 1323 ·
  11. vLLM · IQ4_XS Decode 104 · Prefill 1176 ·
  12. vLLM · Q4_K_M Decode 96 · Prefill 1061 ·
  13. vLLM · Q5_K_M Decode 95 · Prefill 815 ·
  14. Ollama · Q4_K_M Decode 113 · Prefill 922 ·
  15. Ollama · Q5_K_M Decode 95 · Prefill 1090 ·
  16. llama.cpp · Q5_K_M Decode 133.3 · Prefill 999.8 ·

Core figures

Decode (avg)
142.71 tok/s
Prefill (avg)
813.73 tok/s
TTFT (avg)
1.2 s
VRAM (avg)
10.64 GB
MTP acceptance rate
—
TTFB
1 GB
Power draw
302.1 W

Configuration

Member-level fields are taken from the most recent run

Model
Qwen3-8B
Quantization
Q4_K_M
Framework
llama.cpp
Version
v0.48
Context length
32768 tokens
Batch size
512
GPU layers
33
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)
10.64 GB
OS
Windows 11 23H2
Driver
560.94
CUDA
12.8
Power draw
302.1 W

Sources & evidence

2 measured records in total, each traceable to its original source

  1. L2 Cross-framework verifiedOtherOriginal link Verified on 2026-10-19

    v0.48 · Windows 11 23H2 · CUDA 12.8 · 32768 ctx

    124.91 tok/s

    Decode

    423.65 tok/s

    Prefill

    2.32 s

    TTFT

    15.88 GB

    VRAM

    —

    MTP

    302.1 W

    Power draw

    L2 batch seed 2026-10-19: llama.cpp on hw_id=1 running model_id=53 at Q4_K_M
  2. L0 Self-reportedOtherOriginal link Verified on 2026-10-06

    b4200 · Linux · CUDA 12.6 · 4096 ctx

    160.5 tok/s

    Decode

    1,203.8 tok/s

    Prefill

    0.08 s

    TTFT

    5.4 GB

    VRAM

    —

    MTP

    — W

    Power draw

    L0 理论估算(add-1000-records-batch3)