Arao AI
Back to search results

Qwen3-4B · FP16

Start a discussion
NVIDIA RTX 509032GBllama.cpp v0.36L2 Cross-framework verified

This page aggregates 1 real-world runs of Qwen3-4B (FP16) on NVIDIA RTX 5090 with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.

149.3 tok/s

Decode

Decode speed

589.87 tok/s

Prefill

Prefill speed

1.74 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 149.3 · Prefill 589.87 ·
  2. llama.cpp · IQ4_XS Decode 335.8 · Prefill 1082.78 ·
  3. Ollama · IQ3_XS Decode 349.78 · Prefill 1805.86 ·

Core figures

Decode (avg)
149.3 tok/s
Prefill (avg)
589.87 tok/s
TTFT (avg)
1.74 s
VRAM (avg)
17.64 GB
MTP acceptance rate
—
TTFB
1.19 GB
Power draw
271.3 W

Configuration

Member-level fields are taken from the most recent run

Model
Qwen3-4B
Quantization
FP16
Framework
llama.cpp
Version
v0.36
Context length
4096 tokens
Batch size
256
GPU layers
28
Flash Attention
Off

Hardware

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

GPU
NVIDIA RTX 5090
Nominal VRAM
32 GB
Measured VRAM (avg)
17.64 GB
OS
Ubuntu 24.04
Driver
555.85
CUDA
12.6
Power draw
271.3 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

    v0.36 · Ubuntu 24.04 · CUDA 12.6 · 4096 ctx

    149.3 tok/s

    Decode

    589.87 tok/s

    Prefill

    1.74 s

    TTFT

    17.64 GB

    VRAM

    —

    MTP

    271.3 W

    Power draw

    L2 batch seed 2026-10-16: llama.cpp on hw_id=13 running model_id=54 at FP16