Arao AI
Back to search results

Qwen3.5-4B · IQ3_XS

Start a discussion
AMD Ryzen AI Max+ 395128GBllama.cpp v0.31L2 Cross-framework verified

This page aggregates 1 real-world runs of Qwen3.5-4B (IQ3_XS) on AMD Ryzen AI Max+ 395 with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.

156.15 tok/s

Decode

Decode speed

562.48 tok/s

Prefill

Prefill speed

0.48 s

TTFT

Time to first token

7.31 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 · IQ3_XS (current)Decode 156.15 · Prefill 562.48 ·
  2. vLLM · IQ2_XS Decode 205.56 · Prefill 924.99 ·
  3. vLLM · Q6_K Decode 110.75 · Prefill 500.4 ·

Core figures

Decode (avg)
156.15 tok/s
Prefill (avg)
562.48 tok/s
TTFT (avg)
0.48 s
VRAM (avg)
7.31 GB
MTP acceptance rate
—
TTFB
1.25 GB
Power draw
92.4 W

Configuration

Member-level fields are taken from the most recent run

Model
Qwen3.5-4B
Quantization
IQ3_XS
Framework
llama.cpp
Version
v0.31
Context length
16384 tokens
Batch size
512
GPU layers
28
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)
7.31 GB
OS
Ubuntu 22.04
Driver
565.57
CUDA
N/A
Power draw
92.4 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

    v0.31 · Ubuntu 22.04 · CUDA N/A · 16384 ctx

    156.15 tok/s

    Decode

    562.48 tok/s

    Prefill

    0.48 s

    TTFT

    7.31 GB

    VRAM

    —

    MTP

    92.4 W

    Power draw

    L2 batch seed 2026-10-16: llama.cpp on hw_id=24 running model_id=48 at IQ3_XS