Arao AI
Back to search results

Qwen3-4B · FP16

Start a discussion
2x AMD Strix Halollama.cpp v0.42L2 Cross-framework verified

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

87.02 tok/s

Decode

Decode speed

443.84 tok/s

Prefill

Prefill speed

0.62 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 87.02 · Prefill 443.84 ·
  2. Ollama · IQ3_S Decode 197.15 · Prefill 955.03 ·
  3. Ollama · Q4_K_M Decode 165.67 · Prefill 892.04 ·

Core figures

Decode (avg)
87.02 tok/s
Prefill (avg)
443.84 tok/s
TTFT (avg)
0.62 s
VRAM (avg)
17.64 GB
MTP acceptance rate
—
TTFB
0.7 GB
Power draw
83.9 W

Configuration

Member-level fields are taken from the most recent run

Model
Qwen3-4B
Quantization
FP16
Framework
llama.cpp
Version
v0.42
Context length
4096 tokens
Batch size
512
GPU layers
99
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
Windows 11 23H2
Driver
amdgpu 6.11
CUDA
N/A
Power draw
83.9 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

    v0.42 · Windows 11 23H2 · CUDA N/A · 4096 ctx

    87.02 tok/s

    Decode

    443.84 tok/s

    Prefill

    0.62 s

    TTFT

    17.64 GB

    VRAM

    —

    MTP

    83.9 W

    Power draw

    L2 batch seed 2026-10-16: llama.cpp on hw_id=88 running model_id=54 at FP16