Arao AI
Back to search results

Qwen3.5-4B · Q2_K

Start a discussion
NVIDIA DGX Spark128GBvLLM v0.39L2 Cross-framework verified

This page aggregates 1 real-world runs of Qwen3.5-4B (Q2_K) on NVIDIA DGX Spark with vLLM, contributed by 1 independent source platforms; metrics are averages of published measurements.

266.26 tok/s

Decode

Decode speed

1,108.22 tok/s

Prefill

Prefill speed

0.41 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

5 days ago

Last verified

Performance

  1. vLLM · IQ1_S Decode 306.37 · Prefill 1243.41 ·
  2. Ollama · IQ2_XS Decode 258.67 · Prefill 1172.95 ·
  3. vLLM · Q2_K (current)Decode 266.26 · Prefill 1108.22 ·
  4. vLLM · Q3_K_L Decode 219.42 · Prefill 756.56 ·

Core figures

Decode (avg)
266.26 tok/s
Prefill (avg)
1,108.22 tok/s
TTFT (avg)
0.41 s
VRAM (avg)
6.72 GB
MTP acceptance rate
—
TTFB
2.13 GB
Power draw
254 W

Configuration

Member-level fields are taken from the most recent run

Model
Qwen3.5-4B
Quantization
Q2_K
Framework
vLLM
Version
v0.39
Context length
4096 tokens
Batch size
8
GPU layers
999
Flash Attention
Off

Hardware

Nominal and measured figures are shown side by side; whether it runs is the reader's call

GPU
NVIDIA DGX Spark
Nominal VRAM
128 GB
Measured VRAM (avg)
6.72 GB
OS
Ubuntu 24.04
Driver
560.94
CUDA
12.8
Power draw
254 W

Sources & evidence

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

  1. L2 Cross-framework verifiedOtherOriginal link Verified on 2026-10-13

    v0.39 · Ubuntu 24.04 · CUDA 12.8 · 4096 ctx

    266.26 tok/s

    Decode

    1,108.22 tok/s

    Prefill

    0.41 s

    TTFT

    6.72 GB

    VRAM

    —

    MTP

    254 W

    Power draw

    L2 batch seed 2026-10-13: vLLM on hw_id=54 running model_id=48 at Q2_K