Arao AI
Back to search results

Qwen3.5-4B · Q4_K_S

Start a discussion
NVIDIA RTX 509032GBllama.cpp v0.24L2 Cross-framework verified

This page aggregates 2 real-world runs of Qwen3.5-4B (Q4_K_S) on NVIDIA RTX 5090 with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.

334.66 tok/s

Decode

Decode speed

1,225.86 tok/s

Prefill

Prefill speed

1.58 s

TTFT

Time to first token

8.65 GB

VRAM

VRAM usage

L2 Cross-framework verified

2

Measured runs

1

Independent sources

Other

Source platforms

Today

Last verified

Performance

  1. llama.cpp · Q3_K_M Decode 377.37 · Prefill 1718.05 ·
  2. llama.cpp · Q4_K_S (current)Decode 334.66 · Prefill 1225.86 ·
  3. Strata · Q2_K Decode 338.27 · Prefill 1161 ·
  4. Strata · Q4_K_M Decode 318.56 · Prefill 1437.9 ·
  5. llama.cpp · IQ2_M Decode 354.14 · Prefill 1511.04 ·
  6. Ollama · IQ2_M Decode 347.38 · Prefill 1126.85 ·
  7. Ollama · IQ2_XS Decode 333.52 · Prefill 1258.8 ·
  8. llama.cpp · Q3_K_L Decode 338.21 · Prefill 1409.78 ·
  9. vLLM · Q3_K_M Decode 356.1 · Prefill 1633.97 ·
  10. vLLM · Q3_K_S Decode 324.89 · Prefill 1253.12 ·
  11. Strata · Q4_K_S Decode 353.67 · Prefill 1676.75 ·

Core figures

Decode (avg)
334.66 tok/s
Prefill (avg)
1,225.86 tok/s
TTFT (avg)
1.58 s
VRAM (avg)
8.65 GB
MTP acceptance rate
—
TTFB
2.68 GB
Power draw
331.6 W

Configuration

Member-level fields are taken from the most recent run

Model
Qwen3.5-4B
Quantization
Q4_K_S
Framework
llama.cpp
Version
v0.24
Context length
4096 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 5090
Nominal VRAM
32 GB
Measured VRAM (avg)
8.65 GB
OS
Windows 11 23H2
Driver
565.57
CUDA
12.6
Power draw
416.8 W

Sources & evidence

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

  1. L2 Cross-framework verifiedOtherOriginal link Verified on 2026-10-18

    v0.24 · Windows 11 23H2 · CUDA 12.6 · 4096 ctx

    315.71 tok/s

    Decode

    1,124.02 tok/s

    Prefill

    1.57 s

    TTFT

    8.65 GB

    VRAM

    —

    MTP

    416.8 W

    Power draw

    L2 batch seed 2026-10-18: llama.cpp on hw_id=13 running model_id=48 at Q4_K_S
  2. L2 Cross-framework verifiedOtherOriginal link Verified on 2026-10-16

    v0.52 · Ubuntu 22.04 · CUDA 12.6 · 16384 ctx

    353.61 tok/s

    Decode

    1,327.69 tok/s

    Prefill

    1.6 s

    TTFT

    8.65 GB

    VRAM

    —

    MTP

    246.4 W

    Power draw

    L2 batch seed 2026-10-16: llama.cpp on hw_id=13 running model_id=48 at Q4_K_S