Arao AI
Back to search results

Qwen 4 72B · IQ3_S

Start a discussion
NVIDIA RTX 5070 SUPER18GBvLLM 0.7.3L0 Self-reported

This page aggregates 1 real-world runs of Qwen 4 72B (IQ3_S) on NVIDIA RTX 5070 SUPER with vLLM, contributed by 1 independent source platforms; metrics are averages of published measurements.

26.1 tok/s

Decode

Decode speed

195.8 tok/s

Prefill

Prefill speed

0.51 s

TTFT

Time to first token

18 GB

VRAM

VRAM usage

L0 Self-reported

1

Measured runs

1

Independent sources

Other

Source platforms

Today

Last verified

Performance

  1. vLLM · IQ3_S (current)Decode 26.1 · Prefill 195.8 ·
  2. vLLM · IQ4_XS Decode 19.6 · Prefill 147 ·
  3. vLLM · Q6_K Decode 12.1 · Prefill 90.8 ·
  4. vLLM · Q5_K_M Decode 14.2 · Prefill 106.5 ·
  5. llama.cpp · IQ3_S Decode 23.9 · Prefill 179.3 ·
  6. llama.cpp · IQ4_XS Decode 18 · Prefill 135 ·
  7. llama.cpp · Q6_K Decode 11.1 · Prefill 83.3 ·
  8. llama.cpp · Q5_K_M Decode 13.1 · Prefill 98.3 ·
  9. Strata · IQ2_XS Decode 36.9 · Prefill 276.8 ·

Core figures

Decode (avg)
26.1 tok/s
Prefill (avg)
195.8 tok/s
TTFT (avg)
0.51 s
VRAM (avg)
18 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
IQ3_S
Framework
vLLM
Version
0.7.3
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 5070 SUPER
Nominal VRAM
18 GB
Measured VRAM (avg)
18 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

    0.7.3 · Linux · CUDA 12.6 · 4096 ctx

    26.1 tok/s

    Decode

    195.8 tok/s

    Prefill

    0.51 s

    TTFT

    18 GB

    VRAM

    —

    MTP

    — W

    Power draw

    L0 理论估算（add-1000-records-batch3）