Arao AI
Back to search results

gemma-4-E4B-it · IQ3_S

Start a discussion
2x AMD Strix HalovLLM v0.52L2 Cross-framework verified

This page aggregates 1 real-world runs of gemma-4-E4B-it (IQ3_S) on 2x AMD Strix Halo with vLLM, contributed by 1 independent source platforms; metrics are averages of published measurements.

197.35 tok/s

Decode

Decode speed

823.81 tok/s

Prefill

Prefill speed

1.44 s

TTFT

Time to first token

7.64 GB

VRAM

VRAM usage

L2 Cross-framework verified

1

Measured runs

1

Independent sources

Other

Source platforms

2 days ago

Last verified

Performance

  1. vLLM · IQ3_S (current)Decode 197.35 · Prefill 823.81 ·
  2. vLLM · Q5_K_M Decode 149.23 · Prefill 534.97 ·
  3. Strata · Q2_K Decode 311.82 · Prefill 1554.79 ·

Core figures

Decode (avg)
197.35 tok/s
Prefill (avg)
823.81 tok/s
TTFT (avg)
1.44 s
VRAM (avg)
7.64 GB
MTP acceptance rate
—
TTFB
1.58 GB
Power draw
91.4 W

Configuration

Member-level fields are taken from the most recent run

Model
gemma-4-E4B-it
Quantization
IQ3_S
Framework
vLLM
Version
v0.52
Context length
2048 tokens
Batch size
16
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
2x AMD Strix Halo
Measured VRAM (avg)
7.64 GB
OS
macOS 15.0
Driver
565.57
CUDA
N/A
Power draw
91.4 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

    v0.52 · macOS 15.0 · CUDA N/A · 2048 ctx

    197.35 tok/s

    Decode

    823.81 tok/s

    Prefill

    1.44 s

    TTFT

    7.64 GB

    VRAM

    —

    MTP

    91.4 W

    Power draw

    L2 batch seed 2026-10-16: vLLM on hw_id=88 running model_id=67 at IQ3_S