Arao AI
Back to search results

Qwen3-4B-Thinking-2507 · Q4_K_M

Start a discussion
2x AMD Strix Halollama.cpp v0.46L2 Cross-framework verified

This page aggregates 1 real-world runs of Qwen3-4B-Thinking-2507 (Q4_K_M) on 2x AMD Strix Halo with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.

153.87 tok/s

Decode

Decode speed

676.15 tok/s

Prefill

Prefill speed

1.72 s

TTFT

Time to first token

8.99 GB

VRAM

VRAM usage

L2 Cross-framework verified

1

Measured runs

1

Independent sources

Other

Source platforms

2 days ago

Last verified

Performance

  1. llama.cpp · Q3_K_M Decode 181.31 · Prefill 944.53 ·
  2. llama.cpp · Q4_K_M (current)Decode 153.87 · Prefill 676.15 ·
  3. Strata · IQ2_M Decode 312.05 · Prefill 1568.66 ·

Core figures

Decode (avg)
153.87 tok/s
Prefill (avg)
676.15 tok/s
TTFT (avg)
1.72 s
VRAM (avg)
8.99 GB
MTP acceptance rate
—
TTFB
1.23 GB
Power draw
74.8 W

Configuration

Member-level fields are taken from the most recent run

Model
Qwen3-4B-Thinking-2507
Quantization
Q4_K_M
Framework
llama.cpp
Version
v0.46
Context length
8192 tokens
Batch size
256
GPU layers
20
Flash Attention
Off

Hardware

Nominal and measured figures are shown side by side; whether it runs is the reader's call

GPU
2x AMD Strix Halo
Measured VRAM (avg)
8.99 GB
OS
macOS 15.0
Driver
amdgpu 6.11
CUDA
N/A
Power draw
74.8 W

Sources & evidence

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

  1. L2 Cross-framework verifiedOtherOriginal link Verified on 2026-10-16

    v0.46 · macOS 15.0 · CUDA N/A · 8192 ctx

    153.87 tok/s

    Decode

    676.15 tok/s

    Prefill

    1.72 s

    TTFT

    8.99 GB

    VRAM

    —

    MTP

    74.8 W

    Power draw

    L2 batch seed 2026-10-16: llama.cpp on hw_id=88 running model_id=57 at Q4_K_M