Arao AI
Back to search results

gemma-4-12B-it · Q5_K_S

Start a discussion
Apple M3 Max36GBllama.cpp v0.20L2 Cross-framework verified

This page aggregates 1 real-world runs of gemma-4-12B-it (Q5_K_S) on Apple M3 Max with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.

77.34 tok/s

Decode

Decode speed

357.59 tok/s

Prefill

Prefill speed

0.43 s

TTFT

Time to first token

24.78 GB

VRAM

VRAM usage

L2 Cross-framework verified

1

Measured runs

1

Independent sources

Other

Source platforms

Today

Last verified

Performance

  1. llama.cpp · Q4_K_M Decode 85.7 · Prefill 345.14 ·
  2. llama.cpp · Q5_K_S (current)Decode 77.34 · Prefill 357.59 ·
  3. Ollama · IQ3_S Decode 94.58 · Prefill 396.89 ·

Core figures

Decode (avg)
77.34 tok/s
Prefill (avg)
357.59 tok/s
TTFT (avg)
0.43 s
VRAM (avg)
24.78 GB
MTP acceptance rate
—
TTFB
0.62 GB
Power draw
74.5 W

Configuration

Member-level fields are taken from the most recent run

Model
gemma-4-12B-it
Quantization
Q5_K_S
Framework
llama.cpp
Version
v0.20
Context length
16384 tokens
Batch size
512
GPU layers
20
Flash Attention
On

Hardware

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

GPU
Apple M3 Max
Nominal VRAM
36 GB
Measured VRAM (avg)
24.78 GB
OS
macOS 15.0
Driver
Apple M3 Max
CUDA
N/A
Power draw
74.5 W

Sources & evidence

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

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

    v0.20 · macOS 15.0 · CUDA N/A · 16384 ctx

    77.34 tok/s

    Decode

    357.59 tok/s

    Prefill

    0.43 s

    TTFT

    24.78 GB

    VRAM

    —

    MTP

    74.5 W

    Power draw

    L2 batch seed 2026-10-10: llama.cpp on hw_id=4 running model_id=65 at Q5_K_S