Arao AI
Back to search results

Qwen3-4B · Q6_K

Start a discussion
AMD Ryzen AI Max+ 395128GBllama.cpp v0.23L2 Cross-framework verified

This page aggregates 1 real-world runs of Qwen3-4B (Q6_K) on AMD Ryzen AI Max+ 395 with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.

104.54 tok/s

Decode

Decode speed

394.29 tok/s

Prefill

Prefill speed

1.55 s

TTFT

Time to first token

10.5 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 · Q6_K (current)Decode 104.54 · Prefill 394.29 ·
  2. Ollama · IQ1_S Decode 170.91 · Prefill 585.17 ·
  3. Ollama · IQ2_M Decode 170.8 · Prefill 683.91 ·
  4. Ollama · Q2_K Decode 165.94 · Prefill 894.2 ·
  5. Ollama · Q5_K_S Decode 109.99 · Prefill 499.66 ·

Core figures

Decode (avg)
104.54 tok/s
Prefill (avg)
394.29 tok/s
TTFT (avg)
1.55 s
VRAM (avg)
10.5 GB
MTP acceptance rate
—
TTFB
0.84 GB
Power draw
92.9 W

Configuration

Member-level fields are taken from the most recent run

Model
Qwen3-4B
Quantization
Q6_K
Framework
llama.cpp
Version
v0.23
Context length
2048 tokens
Batch size
512
GPU layers
33
Flash Attention
On

Hardware

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

GPU
AMD Ryzen AI Max+ 395
Nominal VRAM
128 GB
Measured VRAM (avg)
10.5 GB
OS
Ubuntu 22.04
Driver
570.86
CUDA
N/A
Power draw
92.9 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

    v0.23 · Ubuntu 22.04 · CUDA N/A · 2048 ctx

    104.54 tok/s

    Decode

    394.29 tok/s

    Prefill

    1.55 s

    TTFT

    10.5 GB

    VRAM

    —

    MTP

    92.9 W

    Power draw

    L2 batch seed 2026-10-16: llama.cpp on hw_id=24 running model_id=54 at Q6_K