Arao AI
Back to search results

Llama-3.2-3B-Instruct · Q3_K_S

Start a discussion
Apple M5 Ultra512GBllama.cpp v0.41L2 Cross-framework verified

This page aggregates 1 real-world runs of Llama-3.2-3B-Instruct (Q3_K_S) on Apple M5 Ultra with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.

338.26 tok/s

Decode

Decode speed

1,583.22 tok/s

Prefill

Prefill speed

0.93 s

TTFT

Time to first token

6.2 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 · Q3_K_S (current)Decode 338.26 · Prefill 1583.22 ·
  2. Ollama · Q3_K_S Decode 318.38 · Prefill 1403.79 ·

Core figures

Decode (avg)
338.26 tok/s
Prefill (avg)
1,583.22 tok/s
TTFT (avg)
0.93 s
VRAM (avg)
6.2 GB
MTP acceptance rate
—
TTFB
2.71 GB
Power draw
123.7 W

Configuration

Member-level fields are taken from the most recent run

Model
Llama-3.2-3B-Instruct
Quantization
Q3_K_S
Framework
llama.cpp
Version
v0.41
Context length
32768 tokens
Batch size
512
GPU layers
99
Flash Attention
On

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)
6.2 GB
OS
macOS 26.1
Power draw
123.7 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

    v0.41 · macOS 26.1 · 32768 ctx

    338.26 tok/s

    Decode

    1,583.22 tok/s

    Prefill

    0.93 s

    TTFT

    6.2 GB

    VRAM

    —

    MTP

    123.7 W

    Power draw

    L2 batch seed 2026-10-19: llama.cpp on hw_id=29 running model_id=70 at Q3_K_S