Arao AI
Back to search results

gemma-4-26B-A4B-it · IQ3_S

Start a discussion
NVIDIA DGX Spark128GBllama.cpp v0.35L2 Cross-framework verified

This page aggregates 1 real-world runs of gemma-4-26B-A4B-it (IQ3_S) on NVIDIA DGX Spark with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.

112.41 tok/s

Decode

Decode speed

545.82 tok/s

Prefill

Prefill speed

1.36 s

TTFT

Time to first token

38.14 GB

VRAM

VRAM usage

L2 Cross-framework verified

1

Measured runs

1

Independent sources

Other

Source platforms

5 days ago

Last verified

Performance

  1. llama.cpp · IQ3_S (current)Decode 112.41 · Prefill 545.82 ·
  2. llama.cpp · Q3_K_L Decode 105.58 · Prefill 348.6 ·
  3. Ollama · Q3_K_L Decode 104.18 · Prefill 516.57 ·
  4. Ollama · Q3_K_M Decode 113.11 · Prefill 567.16 ·

Core figures

Decode (avg)
112.41 tok/s
Prefill (avg)
545.82 tok/s
TTFT (avg)
1.36 s
VRAM (avg)
38.14 GB
MTP acceptance rate
—
TTFB
0.9 GB
Power draw
245 W

Configuration

Member-level fields are taken from the most recent run

Model
gemma-4-26B-A4B-it
Quantization
IQ3_S
Framework
llama.cpp
Version
v0.35
Context length
32768 tokens
Batch size
512
GPU layers
99
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)
38.14 GB
OS
Ubuntu 24.04
Driver
560.94
CUDA
12.8
Power draw
245 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

    v0.35 · Ubuntu 24.04 · CUDA 12.8 · 32768 ctx

    112.41 tok/s

    Decode

    545.82 tok/s

    Prefill

    1.36 s

    TTFT

    38.14 GB

    VRAM

    —

    MTP

    245 W

    Power draw

    L2 batch seed 2026-10-13: llama.cpp on hw_id=54 running model_id=64 at IQ3_S