Arao AI
Back to search results

Llama-3.2-3B-Instruct · IQ3_S

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

This page aggregates 1 real-world runs of Llama-3.2-3B-Instruct (IQ3_S) on NVIDIA RTX 4090 with vLLM, contributed by 1 independent source platforms; metrics are averages of published measurements.

362.35 tok/s

Decode

Decode speed

1,574.76 tok/s

Prefill

Prefill speed

0.64 s

TTFT

Time to first token

6.26 GB

VRAM

VRAM usage

L2 Cross-framework verified

1

Measured runs

1

Independent sources

Other

Source platforms

Today

Last verified

Performance

  1. vLLM · IQ3_S (current)Decode 362.35 · Prefill 1574.76 ·
  2. Strata · IQ1_S Decode 284.1 · Prefill 1178.93 ·
  3. Strata · IQ2_XS Decode 281.21 · Prefill 1389.98 ·
  4. llama.cpp · Q8_0 Decode 245.7 · Prefill 1842.8 ·
  5. llama.cpp · Q5_K_M Decode 357.4 · Prefill 2680.5 ·
  6. llama.cpp · Q4_K_M Decode 436.8 · Prefill 3276 ·

Core figures

Decode (avg)
362.35 tok/s
Prefill (avg)
1,574.76 tok/s
TTFT (avg)
0.64 s
VRAM (avg)
6.26 GB
MTP acceptance rate
—
TTFB
2.9 GB
Power draw
342 W

Configuration

Member-level fields are taken from the most recent run

Model
Llama-3.2-3B-Instruct
Quantization
IQ3_S
Framework
vLLM
Version
v0.47
Context length
32768 tokens
Batch size
32
GPU layers
999
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)
6.26 GB
OS
Ubuntu 24.04
Driver
572.83
CUDA
12.6
Power draw
342 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 · Ubuntu 24.04 · CUDA 12.6 · 32768 ctx

    362.35 tok/s

    Decode

    1,574.76 tok/s

    Prefill

    0.64 s

    TTFT

    6.26 GB

    VRAM

    —

    MTP

    342 W

    Power draw

    L2 batch seed 2026-10-18: vLLM on hw_id=1 running model_id=70 at IQ3_S