Arao AI
Back to search results

Qwen3-4B · Q4_K_S

Start a discussion
AMD Strix Halo (Radeon 8060S)llama.cpp v0.34L2 Cross-framework verified

This page aggregates 1 real-world runs of Qwen3-4B (Q4_K_S) on AMD Strix Halo (Radeon 8060S) with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.

106.73 tok/s

Decode

Decode speed

450.35 tok/s

Prefill

Prefill speed

1.18 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

2 days ago

Last verified

Performance

  1. llama.cpp · Q4_K_S (current)Decode 106.73 · Prefill 450.35 ·
  2. Strata · Q4_K_S Decode 150.62 · Prefill 773.06 ·
  3. Strata · Q6_K Decode 108.81 · Prefill 551.07 ·

Core figures

Decode (avg)
106.73 tok/s
Prefill (avg)
450.35 tok/s
TTFT (avg)
1.18 s
VRAM (avg)
8.65 GB
MTP acceptance rate
—
TTFB
0.85 GB
Power draw
106 W

Configuration

Member-level fields are taken from the most recent run

Model
Qwen3-4B
Quantization
Q4_K_S
Framework
llama.cpp
Version
v0.34
Context length
16384 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
AMD Strix Halo (Radeon 8060S)
Measured VRAM (avg)
8.65 GB
OS
Ubuntu 22.04
Driver
565.57
CUDA
N/A
Power draw
106 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

    v0.34 · Ubuntu 22.04 · CUDA N/A · 16384 ctx

    106.73 tok/s

    Decode

    450.35 tok/s

    Prefill

    1.18 s

    TTFT

    8.65 GB

    VRAM

    —

    MTP

    106 W

    Power draw

    L2 batch seed 2026-10-16: llama.cpp on hw_id=87 running model_id=54 at Q4_K_S