Introducing Shieldstral.

New

Open-weights safety classifier for detecting harmful multimodal content.

AI Security & Compliance

8.2 (61.717 score)

open-source

Visit Tool

Overview

Shieldstral is a 3B parameter open-source safety classifier designed to detect harmful content across text and images. Built by Mistral, it outperforms larger models while remaining computationally efficient. It's built for developers and organizations needing content moderation without proprietary dependencies.

Pros

Outperforms models up to 7x larger in safety classification
Open-weights design enables local deployment without vendor lock-in
Multimodal capability detects harm across text and image inputs
3B parameters keep inference costs and latency low

✕ Cons

Limited documentation on real-world accuracy benchmarks available
Requires technical expertise to integrate into existing systems
No official managed API or hosted inference option

Key Features

Open-source safety classifier

Multimodal content detection

Lightweight 3B model

Local deployment capability

Text and image analysis

Use Cases

Content platforms moderating user-generated text and imagesLLM providers filtering harmful outputs before user deliveryEnterprise applications protecting against toxic or explicit contentDevelopers building moderation systems with local inference