Arao AI
Back to search results

Qwen3.5-4B · Q3_K_S

Start a discussion
Intel Arc B58012GBllama.cpp v0.54L2 Cross-framework verified

This page aggregates 1 real-world runs of Qwen3.5-4B (Q3_K_S) on Intel Arc B580 with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.

74.37 tok/s

Decode

Decode speed

373.85 tok/s

Prefill

Prefill speed

0.91 s

TTFT

Time to first token

7.56 GB

VRAM

VRAM usage

L2 Cross-framework verified

1

Measured runs

1

Independent sources

Other

Source platforms

Today

Last verified

Performance

  1. llama.cpp · Q3_K_S (current)Decode 74.37 · Prefill 373.85 ·
  2. Strata · Q6_K Decode 60.78 · Prefill 199.46 ·
  3. llama.cpp · IQ1_S Decode 98.24 · Prefill 506.66 ·
  4. llama.cpp · Q3_K_L Decode 70.52 · Prefill 317.93 ·
  5. vLLM · Q2_K Decode 89.29 · Prefill 348.31 ·
  6. Strata · IQ3_S Decode 101.32 · Prefill 490.59 ·

Core figures

Decode (avg)
74.37 tok/s
Prefill (avg)
373.85 tok/s
TTFT (avg)
0.91 s
VRAM (avg)
7.56 GB
MTP acceptance rate
—
TTFB
0.59 GB
Power draw
138.9 W

Configuration

Member-level fields are taken from the most recent run

Model
Qwen3.5-4B
Quantization
Q3_K_S
Framework
llama.cpp
Version
v0.54
Context length
2048 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
Intel Arc B580
Nominal VRAM
12 GB
Measured VRAM (avg)
7.56 GB
OS
Windows 11 24H2
Driver
Intel Arc 32.0.101.6552
CUDA
12.6
Power draw
138.9 W

Sources & evidence

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

  1. L2 Cross-framework verifiedOtherOriginal link Verified on 2026-10-18

    v0.54 · Windows 11 24H2 · CUDA 12.6 · 2048 ctx

    74.37 tok/s

    Decode

    373.85 tok/s

    Prefill

    0.91 s

    TTFT

    7.56 GB

    VRAM

    —

    MTP

    138.9 W

    Power draw

    L2 batch seed 2026-10-18: llama.cpp on hw_id=94 running model_id=48 at Q3_K_S