Arao AI
Back to search results

Qwen3-4B-Thinking-2507 · Q5_K_M

Start a discussion
AMD Strix Halo (Radeon 8060S)llama.cpp v0.39L2 Cross-framework verified

This page aggregates 1 real-world runs of Qwen3-4B-Thinking-2507 (Q5_K_M) on AMD Strix Halo (Radeon 8060S) with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.

77.99 tok/s

Decode

Decode speed

423.17 tok/s

Prefill

Prefill speed

0.35 s

TTFT

Time to first token

9.91 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_S Decode 117.04 · Prefill 404.22 ·
  2. llama.cpp · Q5_K_M (current)Decode 77.99 · Prefill 423.17 ·
  3. Strata · IQ2_XS Decode 206.18 · Prefill 1005.37 ·
  4. Strata · IQ3_S Decode 166.26 · Prefill 669.87 ·

Core figures

Decode (avg)
77.99 tok/s
Prefill (avg)
423.17 tok/s
TTFT (avg)
0.35 s
VRAM (avg)
9.91 GB
MTP acceptance rate
—
TTFB
0.62 GB
Power draw
86 W

Configuration

Member-level fields are taken from the most recent run

Model
Qwen3-4B-Thinking-2507
Quantization
Q5_K_M
Framework
llama.cpp
Version
v0.39
Context length
32768 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
AMD Strix Halo (Radeon 8060S)
Measured VRAM (avg)
9.91 GB
OS
Ubuntu 24.04
Driver
565.57
CUDA
N/A
Power draw
86 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

    v0.39 · Ubuntu 24.04 · CUDA N/A · 32768 ctx

    77.99 tok/s

    Decode

    423.17 tok/s

    Prefill

    0.35 s

    TTFT

    9.91 GB

    VRAM

    —

    MTP

    86 W

    Power draw

    L2 batch seed 2026-10-16: llama.cpp on hw_id=87 running model_id=57 at Q5_K_M