Arao AI
Back to search results

gemma-4-12B-it · IQ3_XS

Start a discussion
Apple M5 Ultra512GBllama.cpp v0.38L2 Cross-framework verified

This page aggregates 1 real-world runs of gemma-4-12B-it (IQ3_XS) on Apple M5 Ultra with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.

105.53 tok/s

Decode

Decode speed

434.57 tok/s

Prefill

Prefill speed

1.74 s

TTFT

Time to first token

17.72 GB

VRAM

VRAM usage

L2 Cross-framework verified

1

Measured runs

1

Independent sources

Other

Source platforms

17 days ago

Last verified

Performance

  1. llama.cpp · IQ3_XS (current)Decode 105.53 · Prefill 434.57 ·
  2. Ollama · Q3_K_L Decode 86.92 · Prefill 466.52 ·
  3. Ollama · Q5_K_S Decode 66.58 · Prefill 239.33 ·

Core figures

Decode (avg)
105.53 tok/s
Prefill (avg)
434.57 tok/s
TTFT (avg)
1.74 s
VRAM (avg)
17.72 GB
MTP acceptance rate
—
TTFB
0.84 GB
Power draw
131.6 W

Configuration

Member-level fields are taken from the most recent run

Model
gemma-4-12B-it
Quantization
IQ3_XS
Framework
llama.cpp
Version
v0.38
Context length
8192 tokens
Batch size
512
GPU layers
33
Flash Attention
Off

Hardware

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

GPU
Apple M5 Ultra
Nominal VRAM
512 GB
Measured VRAM (avg)
17.72 GB
OS
macOS 26.1
Power draw
131.6 W

Sources & evidence

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

  1. L2 Cross-framework verifiedOtherOriginal link Verified on 2026-10-19

    v0.38 · macOS 26.1 · 8192 ctx

    105.53 tok/s

    Decode

    434.57 tok/s

    Prefill

    1.74 s

    TTFT

    17.72 GB

    VRAM

    —

    MTP

    131.6 W

    Power draw

    L2 batch seed 2026-10-19: llama.cpp on hw_id=29 running model_id=65 at IQ3_XS