Arao AI
Back to search results

Llama-3.2-3B-Instruct · Q4_K_M

Start a discussion
AMD Ryzen AI Max+ 395128GBvLLM v0.37L2 Cross-framework verified

This page aggregates 1 real-world runs of Llama-3.2-3B-Instruct (Q4_K_M) on AMD Ryzen AI Max+ 395 with vLLM, contributed by 1 independent source platforms; metrics are averages of published measurements.

89.28 tok/s

Decode

Decode speed

327.44 tok/s

Prefill

Prefill speed

2.09 s

TTFT

Time to first token

7.27 GB

VRAM

VRAM usage

L2 Cross-framework verified

1

Measured runs

1

Independent sources

Other

Source platforms

16 days ago

Last verified

Performance

  1. llama.cpp · Q8_0 Decode 61.59 · Prefill 327.72 ·
  2. vLLM · Q4_K_M (current)Decode 89.28 · Prefill 327.44 ·

Core figures

Decode (avg)
89.28 tok/s
Prefill (avg)
327.44 tok/s
TTFT (avg)
2.09 s
VRAM (avg)
7.27 GB
MTP acceptance rate
—
TTFB
0.71 GB
Power draw
74.4 W

Configuration

Member-level fields are taken from the most recent run

Model
Llama-3.2-3B-Instruct
Quantization
Q4_K_M
Framework
vLLM
Version
v0.37
Context length
2048 tokens
Batch size
32
GPU layers
999
Flash Attention
Off

Hardware

Nominal and measured figures are shown side by side; whether it runs is the reader's call

GPU
AMD Ryzen AI Max+ 395
Nominal VRAM
128 GB
Measured VRAM (avg)
7.27 GB
OS
Ubuntu 24.04
Driver
amdgpu 6.12
CUDA
12.6
Power draw
74.4 W

Sources & evidence

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

  1. L2 Cross-framework verifiedOtherOriginal link Verified on 2026-10-19

    v0.37 · Ubuntu 24.04 · CUDA 12.6 · 2048 ctx

    89.28 tok/s

    Decode

    327.44 tok/s

    Prefill

    2.09 s

    TTFT

    7.27 GB

    VRAM

    —

    MTP

    74.4 W

    Power draw

    L2 batch seed 2026-10-19: vLLM on hw_id=24 running model_id=70 at Q4_K_M