Arao AI
Back to search results

Qwen3-8B · Q4_K_M

Start a discussion
NVIDIA DGX Spark128GBllama.cpp v0.54L2 Cross-framework verified

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

146.4 tok/s

Decode

Decode speed

699.64 tok/s

Prefill

Prefill speed

1.05 s

TTFT

Time to first token

15.88 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 · Q3_K_M Decode 167.2 · Prefill 861.79 ·
  2. Strata · Q5_K_S Decode 197.21 · Prefill 856.13 ·
  3. llama.cpp · Q4_K_M (current)Decode 146.4 · Prefill 699.64 ·
  4. vLLM · IQ4_XS Decode 168.28 · Prefill 770.17 ·

Core figures

Decode (avg)
146.4 tok/s
Prefill (avg)
699.64 tok/s
TTFT (avg)
1.05 s
VRAM (avg)
15.88 GB
MTP acceptance rate
—
TTFB
1.17 GB
Power draw
218.9 W

Configuration

Member-level fields are taken from the most recent run

Model
Qwen3-8B
Quantization
Q4_K_M
Framework
llama.cpp
Version
v0.54
Context length
8192 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 DGX Spark
Nominal VRAM
128 GB
Measured VRAM (avg)
15.88 GB
OS
Ubuntu 24.04
Driver
555.85
CUDA
12.8
Power draw
218.9 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

    v0.54 · Ubuntu 24.04 · CUDA 12.8 · 8192 ctx

    146.4 tok/s

    Decode

    699.64 tok/s

    Prefill

    1.05 s

    TTFT

    15.88 GB

    VRAM

    —

    MTP

    218.9 W

    Power draw

    L2 batch seed 2026-10-13: llama.cpp on hw_id=54 running model_id=53 at Q4_K_M