Arao AI
Back to search results

Qwen 4 72B · Q6_K

Start a discussion
NVIDIA RTX 509032GBllama.cpp b4200L0 Self-reported

This page aggregates 1 real-world runs of Qwen 4 72B (Q6_K) on NVIDIA RTX 5090 with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.

22.1 tok/s

Decode

Decode speed

165.8 tok/s

Prefill

Prefill speed

0.6 s

TTFT

Time to first token

32 GB

VRAM

VRAM usage

L0 Self-reported

1

Measured runs

1

Independent sources

Other

Source platforms

Today

Last verified

Performance

  1. vLLM · IQ3_S Decode 52.2 · Prefill 391.5 ·
  2. vLLM · IQ4_XS Decode 39.2 · Prefill 294 ·
  3. vLLM · Q6_K Decode 24.1 · Prefill 180.8 ·
  4. vLLM · Q5_K_M Decode 28.5 · Prefill 213.8 ·
  5. llama.cpp · IQ3_S Decode 47.9 · Prefill 359.3 ·
  6. llama.cpp · IQ4_XS Decode 35.9 · Prefill 269.3 ·
  7. llama.cpp · Q6_K (current)Decode 22.1 · Prefill 165.8 ·
  8. llama.cpp · Q5_K_M Decode 26.1 · Prefill 195.8 ·
  9. vLLM · Q4_K_M Decode 34.9 · Prefill 261.8 ·

Core figures

Decode (avg)
22.1 tok/s
Prefill (avg)
165.8 tok/s
TTFT (avg)
0.6 s
VRAM (avg)
32 GB
MTP acceptance rate
—
TTFB
— GB
Power draw
— W

Configuration

Member-level fields are taken from the most recent run

Model
Qwen 4 72B
Quantization
Q6_K
Framework
llama.cpp
Version
b4200
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
NVIDIA RTX 5090
Nominal VRAM
32 GB
Measured VRAM (avg)
32 GB
OS
Linux
Driver
570.65
CUDA
12.6

Sources & evidence

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

  1. L0 Self-reportedOtherOriginal link Verified on 2026-10-06

    b4200 · Linux · CUDA 12.6 · 4096 ctx

    22.1 tok/s

    Decode

    165.8 tok/s

    Prefill

    0.6 s

    TTFT

    32 GB

    VRAM

    —

    MTP

    — W

    Power draw

    L0 理论估算（add-1000-records-batch3）