Arao AI
Back to search results

Qwen3.5-4B · FP16

Start a discussion
NVIDIA RTX 409024GBllama.cpp v0.47L2 Cross-framework verified

This page aggregates 1 real-world runs of Qwen3.5-4B (FP16) on NVIDIA RTX 4090 with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.

79.5 tok/s

Decode

Decode speed

355.19 tok/s

Prefill

Prefill speed

1.85 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

Today

Last verified

Performance

  1. llama.cpp · FP16 (current)Decode 79.5 · Prefill 355.19 ·
  2. llama.cpp · IQ4_XS Decode 170.35 · Prefill 636.3 ·
  3. vLLM · IQ2_M Decode 248.85 · Prefill 1225.63 ·
  4. vLLM · Q6_K Decode 133.61 · Prefill 629.89 ·
  5. llama.cpp · Q3_K_L Decode 174.81 · Prefill 809.7 ·
  6. vLLM · Q4_K_S Decode 181.53 · Prefill 809.47 ·
  7. Strata · Q6_K Decode 161.44 · Prefill 654.89 ·

Core figures

Decode (avg)
79.5 tok/s
Prefill (avg)
355.19 tok/s
TTFT (avg)
1.85 s
VRAM (avg)
17.64 GB
MTP acceptance rate
—
TTFB
0.64 GB
Power draw
327.3 W

Configuration

Member-level fields are taken from the most recent run

Model
Qwen3.5-4B
Quantization
FP16
Framework
llama.cpp
Version
v0.47
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
NVIDIA RTX 4090
Nominal VRAM
24 GB
Measured VRAM (avg)
17.64 GB
OS
Windows 11 24H2
Driver
560.94
CUDA
12.4
Power draw
327.3 W

Sources & evidence

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

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

    v0.47 · Windows 11 24H2 · CUDA 12.4 · 2048 ctx

    79.5 tok/s

    Decode

    355.19 tok/s

    Prefill

    1.85 s

    TTFT

    17.64 GB

    VRAM

    —

    MTP

    327.3 W

    Power draw

    L2 batch seed 2026-10-18: llama.cpp on hw_id=1 running model_id=48 at FP16