Arao AI
Back to search results

Ministral-3-8B-Instruct-2512 · IQ1_S

Start a discussion
NVIDIA RTX 509032GBllama.cpp v0.37L2 Cross-framework verified

This page aggregates 1 real-world runs of Ministral-3-8B-Instruct-2512 (IQ1_S) on NVIDIA RTX 5090 with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.

381.92 tok/s

Decode

Decode speed

1,480.95 tok/s

Prefill

Prefill speed

0.75 s

TTFT

Time to first token

9.16 GB

VRAM

VRAM usage

L2 Cross-framework verified

1

Measured runs

1

Independent sources

Other

Source platforms

11 days ago

Last verified

Performance

  1. llama.cpp · IQ1_S (current)Decode 381.92 · Prefill 1480.95 ·
  2. vLLM · Q3_K_S Decode 337.39 · Prefill 1195.07 ·

Core figures

Decode (avg)
381.92 tok/s
Prefill (avg)
1,480.95 tok/s
TTFT (avg)
0.75 s
VRAM (avg)
9.16 GB
MTP acceptance rate
—
TTFB
3.06 GB
Power draw
462.4 W

Configuration

Member-level fields are taken from the most recent run

Model
Ministral-3-8B-Instruct-2512
Quantization
IQ1_S
Framework
llama.cpp
Version
v0.37
Context length
65536 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 5090
Nominal VRAM
32 GB
Measured VRAM (avg)
9.16 GB
OS
Ubuntu 24.04
Driver
572.83
CUDA
12.6
Power draw
462.4 W

Sources & evidence

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

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

    v0.37 · Ubuntu 24.04 · CUDA 12.6 · 65536 ctx

    381.92 tok/s

    Decode

    1,480.95 tok/s

    Prefill

    0.75 s

    TTFT

    9.16 GB

    VRAM

    —

    MTP

    462.4 W

    Power draw

    L2 batch seed 2026-10-24: llama.cpp on hw_id=13 running model_id=74 at IQ1_S