Arao AI
Back to search results

Llama-3.2-3B-Instruct · IQ4_XS

Start a discussion
AMD Radeon RX 907016GBvLLM v0.26L2 Cross-framework verified

This page aggregates 1 real-world runs of Llama-3.2-3B-Instruct (IQ4_XS) on AMD Radeon RX 9070 with vLLM, contributed by 1 independent source platforms; metrics are averages of published measurements.

366.03 tok/s

Decode

Decode speed

1,680.79 tok/s

Prefill

Prefill speed

0.89 s

TTFT

Time to first token

6.76 GB

VRAM

VRAM usage

L2 Cross-framework verified

1

Measured runs

1

Independent sources

Other

Source platforms

Today

Last verified

Performance

  1. llama.cpp · Q4_K_M Decode 313.76 · Prefill 1093.15 ·
  2. llama.cpp · Q4_K_S Decode 290.99 · Prefill 1343.85 ·
  3. vLLM · IQ4_XS (current)Decode 366.03 · Prefill 1680.79 ·
  4. vLLM · Q3_K_S Decode 348.42 · Prefill 1407.74 ·

Core figures

Decode (avg)
366.03 tok/s
Prefill (avg)
1,680.79 tok/s
TTFT (avg)
0.89 s
VRAM (avg)
6.76 GB
MTP acceptance rate
—
TTFB
2.93 GB
Power draw
252.6 W

Configuration

Member-level fields are taken from the most recent run

Model
Llama-3.2-3B-Instruct
Quantization
IQ4_XS
Framework
vLLM
Version
v0.26
Context length
16384 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
AMD Radeon RX 9070
Nominal VRAM
16 GB
Measured VRAM (avg)
6.76 GB
OS
Windows 11 23H2
Driver
amdgpu 6.11
CUDA
12.8
Power draw
252.6 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

    v0.26 · Windows 11 23H2 · CUDA 12.8 · 16384 ctx

    366.03 tok/s

    Decode

    1,680.79 tok/s

    Prefill

    0.89 s

    TTFT

    6.76 GB

    VRAM

    —

    MTP

    252.6 W

    Power draw

    L2 batch seed 2026-10-18: vLLM on hw_id=21 running model_id=70 at IQ4_XS