Arao AI
Back to search results

gemma-4-31B-it · Q3_K_L

Start a discussion
Apple M5 Ultra512GBllama.cpp v0.43L2 Cross-framework verified

This page aggregates 1 real-world runs of gemma-4-31B-it (Q3_K_L) on Apple M5 Ultra with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.

65.89 tok/s

Decode

Decode speed

258.39 tok/s

Prefill

Prefill speed

2.06 s

TTFT

Time to first token

48.97 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 · Q3_K_L (current)Decode 65.89 · Prefill 258.39 ·
  2. llama.cpp · Q8_0 Decode 37.99 · Prefill 165.65 ·
  3. Ollama · Q2_K Decode 75.92 · Prefill 280.65 ·
  4. Ollama · Q8_0 Decode 38.85 · Prefill 140.28 ·

Core figures

Decode (avg)
65.89 tok/s
Prefill (avg)
258.39 tok/s
TTFT (avg)
2.06 s
VRAM (avg)
48.97 GB
MTP acceptance rate
—
TTFB
0.53 GB
Power draw
122.9 W

Configuration

Member-level fields are taken from the most recent run

Model
gemma-4-31B-it
Quantization
Q3_K_L
Framework
llama.cpp
Version
v0.43
Context length
32768 tokens
Batch size
256
GPU layers
28
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)
48.97 GB
OS
macOS 26.1
Power draw
122.9 W

Sources & evidence

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

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

    v0.43 · macOS 26.1 · 32768 ctx

    65.89 tok/s

    Decode

    258.39 tok/s

    Prefill

    2.06 s

    TTFT

    48.97 GB

    VRAM

    —

    MTP

    122.9 W

    Power draw

    L2 batch seed 2026-10-19: llama.cpp on hw_id=29 running model_id=63 at Q3_K_L