Mistral AI Releases Shieldstral 1.0 3B: An Open-Weights Policy-Adaptive Multimodal Safety Classifier Matching Models 7× Its Size
Mistral AI unveiled Shieldstral 1.0 3B, a 3-billion-parameter open-weights safety classifier that dynamically adjusts to user-defined policies at inference time.
- Shieldstral 1.0 3B is an open-weights multimodal safety classifier with 3 billion parameters.
- It adapts to user-defined policies at inference time without requiring retraining.
- The model matches larger models like GPT-OSS-Safeguard-20B on text safety benchmarks.
- Trained on 54.1 million samples, it achieves 84.9% average F1 on text safety and 83.8% on multimodal safety.
Mistral AI has introduced Shieldstral 1.0 3B, a groundbreaking open-weights multimodal safety classifier designed to redefine content moderation. Unlike traditional models that rely on fixed harm taxonomies, Shieldstral frames moderation as a single yes/no question, allowing operators to supply a plain-language policy at inference time. This approach eliminates the need for retraining when policies change, delivering a calibrated safety score from a single forward pass.
Built on the Ministral-3-3B-Base-2512 architecture with a Pixtral vision encoder, Shieldstral 1.0 3B was trained on approximately 54.1 million samples. The model achieves 84.9% average F1 on text safety benchmarks, matching the performance of GPT-OSS-Safeguard-20B, while also delivering 83.8% F1 on multimodal safety and 91.3% on unspecified additional metrics. Its adaptability and efficiency position it as a practical solution for real-time, policy-driven content moderation across text and vision modalities.
Provides a lightweight, adaptable tool for building customizable safety classifiers without retraining overhead.
Enables cost-effective, policy-compliant content moderation systems that can evolve with regulatory or platform requirements.
Highlights Mistral AI's innovation in open-weights AI tools, potentially expanding its market share in safety-critical applications.
- open-weights
- AI models whose weights are publicly available, allowing users to inspect, modify, and deploy them without restrictions.
- policy-adaptive
- A model that adjusts its behavior based on user-defined guidelines or rules provided at inference time.
AI ToolsBackflip AI turns 3D scans into editable CAD models in minutes instead of hours
AI ToolsAngular WebMCP — Your App is Now an AI Tool 🔥🚀
AI ToolsxAI's Imagine Image 2.0 lands just behind OpenAI's GPT-Image-2 in Arena benchmarks
AI ToolsAI Writes 41% of Code. Only 29% of Devs Trust It. Review It Like a Senior Engineer
AI ToolsTencent Cloud Open-Sources TencentDB Agent Memory v2.0: A Team-Level Memory Hub for AI Coding Agents
Iran wins medals at International Olympiad in Artificial Intelligence - Tehran Times
Iranian students won medals at the International Olympiad in Artificial Intelligence, highlighting the country's growing AI education and research capabilities.
RoboticsThe first self-driving vehicle on Mars has proven to be a smashing success
NASA's Perseverance rover has autonomously driven nearly 90% of its total distance on Mars, demonstrating a major leap in planetary robotic autonomy.
AI ResearchFields Medalist who published a paper on AI-driven human extinction now works for OpenAI
Fields Medalist Jacob Tsimerman is leaving academia to join OpenAI, where he will work on AI safety after publishing a paper on AI-driven human extinction risks.
AI ResearchDeepMind’s hurricane breakthrough has surprised weather scientists
DeepMind’s new open-source WeatherNext model delivers accurate hurricane predictions using lower-resolution weather data, giving forecasters an extra day of lead time.
Forget the Artificial Intelligence (AI) Capex Bubble: SpaceX CEO Elon Musk Says the Memory Shortage Will Drive Costs Even Higher - The Motley Fool
Elon Musk states that the memory shortage will drive AI costs higher. He believes this will surpass the current AI capex bubble.
HardwareFirebird Launches CIS Region’s Largest AI Factory in Armenia
Firebird launched the largest AI factory in the CIS region, located in Armenia, powered by NVIDIA's Blackwell and Rubin platforms alongside Dell Technologies' infrastructure.