Arao AI
Back to search results

Qwen3.5-4B · IQ3_S

Start a discussion
AMD Radeon RX 907016GBvLLM v0.40L2 Cross-framework verified

This page aggregates 1 real-world runs of Qwen3.5-4B (IQ3_S) on AMD Radeon RX 9070 with vLLM, contributed by 1 independent source platforms; metrics are averages of published measurements.

146.59 tok/s

Decode

Decode speed

752.13 tok/s

Prefill

Prefill speed

0.15 s

TTFT

Time to first token

7.64 GB

VRAM

VRAM usage

L2 Cross-framework verified

1

Measured runs

1

Independent sources

Other

Source platforms

5 days ago

Last verified

Performance

  1. llama.cpp · IQ2_XS Decode 204.13 · Prefill 1113.36 ·
  2. llama.cpp · Q3_K_L Decode 162.78 · Prefill 836.11 ·
  3. vLLM · Q3_K_S Decode 194.22 · Prefill 736.68 ·
  4. vLLM · Q4_K_M Decode 137.14 · Prefill 684.52 ·
  5. llama.cpp · IQ4_XS Decode 127.63 · Prefill 638.8 ·
  6. vLLM · IQ1_S Decode 172.53 · Prefill 556.88 ·
  7. vLLM · IQ3_S (current)Decode 146.59 · Prefill 752.13 ·

Core figures

Decode (avg)
146.59 tok/s
Prefill (avg)
752.13 tok/s
TTFT (avg)
0.15 s
VRAM (avg)
7.64 GB
MTP acceptance rate
—
TTFB
1.17 GB
Power draw
236.6 W

Configuration

Member-level fields are taken from the most recent run

Model
Qwen3.5-4B
Quantization
IQ3_S
Framework
vLLM
Version
v0.40
Context length
4096 tokens
Batch size
32
GPU layers
999
Flash Attention
On

Hardware

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

GPU
AMD Radeon RX 9070
Nominal VRAM
16 GB
Measured VRAM (avg)
7.64 GB
OS
Ubuntu 24.04
Driver
amdgpu 6.10
CUDA
N/A
Power draw
236.6 W

Sources & evidence

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

  1. L2 Cross-framework verifiedOtherOriginal link Verified on 2026-10-13

    v0.40 · Ubuntu 24.04 · CUDA N/A · 4096 ctx

    146.59 tok/s

    Decode

    752.13 tok/s

    Prefill

    0.15 s

    TTFT

    7.64 GB

    VRAM

    —

    MTP

    236.6 W

    Power draw

    L2 batch seed 2026-10-13: vLLM on hw_id=21 running model_id=48 at IQ3_S