Arao AI
Back to search results

Qwen3-4B-Thinking-2507 · Q4_K_S

Start a discussion
Apple M4 Pro64GBllama.cpp v0.25L2 Cross-framework verified

This page aggregates 1 real-world runs of Qwen3-4B-Thinking-2507 (Q4_K_S) on Apple M4 Pro with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.

123.54 tok/s

Decode

Decode speed

441.3 tok/s

Prefill

Prefill speed

2.25 s

TTFT

Time to first token

8.65 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 · IQ4_XS Decode 138.37 · Prefill 661.86 ·
  2. llama.cpp · Q4_K_S (current)Decode 123.54 · Prefill 441.3 ·
  3. Strata · IQ3_XS Decode 218.39 · Prefill 1145.06 ·
  4. Strata · Q3_K_L Decode 206.28 · Prefill 783.16 ·

Core figures

Decode (avg)
123.54 tok/s
Prefill (avg)
441.3 tok/s
TTFT (avg)
2.25 s
VRAM (avg)
8.65 GB
MTP acceptance rate
—
TTFB
0.99 GB
Power draw
78.5 W

Configuration

Member-level fields are taken from the most recent run

Model
Qwen3-4B-Thinking-2507
Quantization
Q4_K_S
Framework
llama.cpp
Version
v0.25
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 M4 Pro
Nominal VRAM
64 GB
Measured VRAM (avg)
8.65 GB
OS
macOS 15.0
Driver
Apple M4 Pro
CUDA
N/A
Power draw
78.5 W

Sources & evidence

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

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

    v0.25 · macOS 15.0 · CUDA N/A · 32768 ctx

    123.54 tok/s

    Decode

    441.3 tok/s

    Prefill

    2.25 s

    TTFT

    8.65 GB

    VRAM

    —

    MTP

    78.5 W

    Power draw

    L2 batch seed 2026-10-13: llama.cpp on hw_id=25 running model_id=57 at Q4_K_S