Arao AI
Back to search results

Qwen3.5-4B · FP16

Start a discussion
2x AMD Strix Halollama.cpp v0.50L2 Cross-framework verified

This page aggregates 1 real-world runs of Qwen3.5-4B (FP16) on 2x AMD Strix Halo with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.

94.9 tok/s

Decode

Decode speed

326.17 tok/s

Prefill

Prefill speed

2.39 s

TTFT

Time to first token

17.64 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 · FP16 (current)Decode 94.9 · Prefill 326.17 ·
  2. Strata · Q5_K_M Decode 206.52 · Prefill 718.15 ·

Core figures

Decode (avg)
94.9 tok/s
Prefill (avg)
326.17 tok/s
TTFT (avg)
2.39 s
VRAM (avg)
17.64 GB
MTP acceptance rate
—
TTFB
0.76 GB
Power draw
91.3 W

Configuration

Member-level fields are taken from the most recent run

Model
Qwen3.5-4B
Quantization
FP16
Framework
llama.cpp
Version
v0.50
Context length
8192 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
2x AMD Strix Halo
Measured VRAM (avg)
17.64 GB
OS
Ubuntu 22.04
Driver
570.86
CUDA
N/A
Power draw
91.3 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

    v0.50 · Ubuntu 22.04 · CUDA N/A · 8192 ctx

    94.9 tok/s

    Decode

    326.17 tok/s

    Prefill

    2.39 s

    TTFT

    17.64 GB

    VRAM

    —

    MTP

    91.3 W

    Power draw

    L2 batch seed 2026-10-16: llama.cpp on hw_id=88 running model_id=48 at FP16