Arao AI
Back to search results

Qwen2.5-Omni-3B · Q6_K

Start a discussion
NVIDIA RTX 509032GBllama.cpp v0.28L2 Cross-framework verified

This page aggregates 1 real-world runs of Qwen2.5-Omni-3B (Q6_K) on NVIDIA RTX 5090 with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.

277.82 tok/s

Decode

Decode speed

934.6 tok/s

Prefill

Prefill speed

0.26 s

TTFT

Time to first token

8.4 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 · IQ4_XS Decode 375.39 · Prefill 1839.32 ·
  2. llama.cpp · Q6_K (current)Decode 277.82 · Prefill 934.6 ·
  3. Strata · FP16 Decode 238.76 · Prefill 798.02 ·
  4. Strata · IQ3_S Decode 359.64 · Prefill 1323.17 ·
  5. llama.cpp · Q8_0 Decode 436.8 · Prefill 3276 ·
  6. llama.cpp · Q5_K_M Decode 635.3 · Prefill 4764.8 ·
  7. llama.cpp · Q4_K_M Decode 776.5 · Prefill 5823.8 ·

Core figures

Decode (avg)
277.82 tok/s
Prefill (avg)
934.6 tok/s
TTFT (avg)
0.26 s
VRAM (avg)
8.4 GB
MTP acceptance rate
—
TTFB
2.22 GB
Power draw
462.7 W

Configuration

Member-level fields are taken from the most recent run

Model
Qwen2.5-Omni-3B
Quantization
Q6_K
Framework
llama.cpp
Version
v0.28
Context length
2048 tokens
Batch size
512
GPU layers
20
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)
8.4 GB
OS
Ubuntu 22.04
Driver
565.57
CUDA
12.8
Power draw
462.7 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

    v0.28 · Ubuntu 22.04 · CUDA 12.8 · 2048 ctx

    277.82 tok/s

    Decode

    934.6 tok/s

    Prefill

    0.26 s

    TTFT

    8.4 GB

    VRAM

    —

    MTP

    462.7 W

    Power draw

    L2 batch seed 2026-10-16: llama.cpp on hw_id=13 running model_id=60 at Q6_K