Arao AI
Back to search results

Qwen3-4B · Q3_K_L

Start a discussion
Apple M5 Pro64GBllama.cpp v0.52L2 Cross-framework verified

This page aggregates 1 real-world runs of Qwen3-4B (Q3_K_L) on Apple M5 Pro with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.

189.13 tok/s

Decode

Decode speed

672.33 tok/s

Prefill

Prefill speed

0.84 s

TTFT

Time to first token

8.15 GB

VRAM

VRAM usage

L2 Cross-framework verified

1

Measured runs

1

Independent sources

Other

Source platforms

5 days ago

Last verified

Performance

  1. llama.cpp · Q3_K_L (current)Decode 189.13 · Prefill 672.33 ·
  2. llama.cpp · Q3_K_M Decode 187.03 · Prefill 648.58 ·

Core figures

Decode (avg)
189.13 tok/s
Prefill (avg)
672.33 tok/s
TTFT (avg)
0.84 s
VRAM (avg)
8.15 GB
MTP acceptance rate
—
TTFB
1.51 GB
Power draw
98.3 W

Configuration

Member-level fields are taken from the most recent run

Model
Qwen3-4B
Quantization
Q3_K_L
Framework
llama.cpp
Version
v0.52
Context length
16384 tokens
Batch size
256
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 Pro
Nominal VRAM
64 GB
Measured VRAM (avg)
8.15 GB
OS
macOS 15.0
Driver
Apple M5 Pro
CUDA
N/A
Power draw
98.3 W

Sources & evidence

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

  1. L2 Cross-framework verifiedOtherOriginal link Verified on 2026-10-13

    v0.52 · macOS 15.0 · CUDA N/A · 16384 ctx

    189.13 tok/s

    Decode

    672.33 tok/s

    Prefill

    0.84 s

    TTFT

    8.15 GB

    VRAM

    —

    MTP

    98.3 W

    Power draw

    L2 batch seed 2026-10-13: llama.cpp on hw_id=27 running model_id=54 at Q3_K_L