Arao AI
Back to search results

Qwen3-4B · Q3_K_M

Start a discussion
NVIDIA DGX Spark128GBllama.cpp v0.45L2 Cross-framework verified

This page aggregates 1 real-world runs of Qwen3-4B (Q3_K_M) on NVIDIA DGX Spark with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.

206.5 tok/s

Decode

Decode speed

1,076.24 tok/s

Prefill

Prefill speed

0.36 s

TTFT

Time to first token

7.81 GB

VRAM

VRAM usage

L2 Cross-framework verified

1

Measured runs

1

Independent sources

Other

Source platforms

2 days ago

Last verified

Performance

  1. llama.cpp · Q3_K_M (current)Decode 206.5 · Prefill 1076.24 ·
  2. Strata · Q3_K_S Decode 338.04 · Prefill 1544.38 ·
  3. Strata · Q5_K_M Decode 221.01 · Prefill 864.66 ·
  4. vLLM · Q3_K_L Decode 202.71 · Prefill 852.86 ·
  5. Strata · IQ1_S Decode 323.83 · Prefill 1439.4 ·

Core figures

Decode (avg)
206.5 tok/s
Prefill (avg)
1,076.24 tok/s
TTFT (avg)
0.36 s
VRAM (avg)
7.81 GB
MTP acceptance rate
—
TTFB
1.65 GB
Power draw
165.5 W

Configuration

Member-level fields are taken from the most recent run

Model
Qwen3-4B
Quantization
Q3_K_M
Framework
llama.cpp
Version
v0.45
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 DGX Spark
Nominal VRAM
128 GB
Measured VRAM (avg)
7.81 GB
OS
Ubuntu 22.04
Driver
555.85
CUDA
12.4
Power draw
165.5 W

Sources & evidence

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

  1. L2 Cross-framework verifiedOtherOriginal link Verified on 2026-10-16

    v0.45 · Ubuntu 22.04 · CUDA 12.4 · 4096 ctx

    206.5 tok/s

    Decode

    1,076.24 tok/s

    Prefill

    0.36 s

    TTFT

    7.81 GB

    VRAM

    —

    MTP

    165.5 W

    Power draw

    L2 batch seed 2026-10-16: llama.cpp on hw_id=54 running model_id=54 at Q3_K_M