Arao AI
Back to search results

Qwen3.5-4B · Q6_K

Start a discussion
NVIDIA RTX 309024GBllama.cpp v0.52L2 Cross-framework verified

This page aggregates 1 real-world runs of Qwen3.5-4B (Q6_K) on NVIDIA RTX 3090 with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.

108.41 tok/s

Decode

Decode speed

420.36 tok/s

Prefill

Prefill speed

0.9 s

TTFT

Time to first token

10.5 GB

VRAM

VRAM usage

L2 Cross-framework verified

1

Measured runs

1

Independent sources

Other

Source platforms

Today

Last verified

Performance

  1. llama.cpp · Q5_K_S Decode 139.06 · Prefill 548.16 ·
  2. llama.cpp · Q6_K (current)Decode 108.41 · Prefill 420.36 ·
  3. vLLM · Q5_K_S Decode 129.61 · Prefill 696.02 ·

Core figures

Decode (avg)
108.41 tok/s
Prefill (avg)
420.36 tok/s
TTFT (avg)
0.9 s
VRAM (avg)
10.5 GB
MTP acceptance rate
—
TTFB
0.87 GB
Power draw
328.2 W

Configuration

Member-level fields are taken from the most recent run

Model
Qwen3.5-4B
Quantization
Q6_K
Framework
llama.cpp
Version
v0.52
Context length
32768 tokens
Batch size
256
GPU layers
33
Flash Attention
On

Hardware

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

GPU
NVIDIA RTX 3090
Nominal VRAM
24 GB
Measured VRAM (avg)
10.5 GB
OS
macOS 15.0
Driver
565.57
CUDA
12.6
Power draw
328.2 W

Sources & evidence

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

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

    v0.52 · macOS 15.0 · CUDA 12.6 · 32768 ctx

    108.41 tok/s

    Decode

    420.36 tok/s

    Prefill

    0.9 s

    TTFT

    10.5 GB

    VRAM

    —

    MTP

    328.2 W

    Power draw

    L2 batch seed 2026-10-10: llama.cpp on hw_id=2 running model_id=48 at Q6_K