Arao AI
Back to search results

Ministral-3-8B-Instruct-2512 · Q5_K_S

Start a discussion
NVIDIA H2097GBvLLM v0.53L2 Cross-framework verified

This page aggregates 1 real-world runs of Ministral-3-8B-Instruct-2512 (Q5_K_S) on NVIDIA H20 with vLLM, contributed by 1 independent source platforms; metrics are averages of published measurements.

49.64 tok/s

Decode

Decode speed

220.24 tok/s

Prefill

Prefill speed

1.95 s

TTFT

Time to first token

17.22 GB

VRAM

VRAM usage

L2 Cross-framework verified

1

Measured runs

1

Independent sources

Other

Source platforms

12 days ago

Last verified

Performance

  1. vLLM · Q5_K_S (current)Decode 49.64 · Prefill 220.24 ·
  2. Strata · Q4_K_M Decode 73.54 · Prefill 286.94 ·

Core figures

Decode (avg)
49.64 tok/s
Prefill (avg)
220.24 tok/s
TTFT (avg)
1.95 s
VRAM (avg)
17.22 GB
MTP acceptance rate
—
TTFB
0.4 GB
Power draw
331.6 W

Configuration

Member-level fields are taken from the most recent run

Model
Ministral-3-8B-Instruct-2512
Quantization
Q5_K_S
Framework
vLLM
Version
v0.53
Context length
16384 tokens
Batch size
8
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 H20
Nominal VRAM
97 GB
Measured VRAM (avg)
17.22 GB
OS
Ubuntu 24.04
Driver
560.94
CUDA
12.4
Power draw
331.6 W

Sources & evidence

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

  1. L2 Cross-framework verifiedOtherOriginal link Verified on 2026-10-24

    v0.53 · Ubuntu 24.04 · CUDA 12.4 · 16384 ctx

    49.64 tok/s

    Decode

    220.24 tok/s

    Prefill

    1.95 s

    TTFT

    17.22 GB

    VRAM

    —

    MTP

    331.6 W

    Power draw

    L2 batch seed 2026-10-24: vLLM on hw_id=56 running model_id=74 at Q5_K_S