Arao AI
Back to search results

Llama-3.2-3B-Instruct · Q4_K_S

Start a discussion
Apple M5 Max128GBllama.cpp v0.50L2 Cross-framework verified

This page aggregates 1 real-world runs of Llama-3.2-3B-Instruct (Q4_K_S) on Apple M5 Max with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.

188.34 tok/s

Decode

Decode speed

664.86 tok/s

Prefill

Prefill speed

2.23 s

TTFT

Time to first token

7.01 GB

VRAM

VRAM usage

L2 Cross-framework verified

1

Measured runs

1

Independent sources

Other

Source platforms

16 days ago

Last verified

Performance

  1. llama.cpp · Q4_K_S (current)Decode 188.34 · Prefill 664.86 ·
  2. Ollama · IQ3_XS Decode 209.81 · Prefill 739.8 ·

Core figures

Decode (avg)
188.34 tok/s
Prefill (avg)
664.86 tok/s
TTFT (avg)
2.23 s
VRAM (avg)
7.01 GB
MTP acceptance rate
—
TTFB
1.51 GB
Power draw
52.4 W

Configuration

Member-level fields are taken from the most recent run

Model
Llama-3.2-3B-Instruct
Quantization
Q4_K_S
Framework
llama.cpp
Version
v0.50
Context length
8192 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 M5 Max
Nominal VRAM
128 GB
Measured VRAM (avg)
7.01 GB
OS
macOS 26.1
Power draw
52.4 W

Sources & evidence

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

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

    v0.50 · macOS 26.1 · 8192 ctx

    188.34 tok/s

    Decode

    664.86 tok/s

    Prefill

    2.23 s

    TTFT

    7.01 GB

    VRAM

    —

    MTP

    52.4 W

    Power draw

    L2 batch seed 2026-10-19: llama.cpp on hw_id=28 running model_id=70 at Q4_K_S