Top AI Product

Every guard model has the same flaw: the safety taxonomy is baked in at training time. Want different rules? Retrain. Mistral Shieldstral, released August 4 under Apache 2.0, kills that constraint.

Write your policy in plain English

Shieldstral is a 3B open-weights safety classifier. You hand it a moderation policy as a natural-language question, it returns a calibrated safety score — one yes/no answer, one forward pass. No fixed categories, no fine-tuning. It handles text and images, and runs on a single 16GB GPU.

The sharp number: Mistral claims it matches or beats open guard models up to 7x its size across text safety, refusal detection, policy adaptability, and multimodal benchmarks. HackerNews gave it 216 points in a day.

Weights and API access

Weights are free on Hugging Face, with API access through Mistral’s platform. Typical setup: sit it between your agent and the world — score every user input, tool output, and model response against your own policy. That’s the real bet here. Agents generate way more content than humans can review, and a cheap, policy-flexible moderation layer is infrastructure everyone suddenly needs.

You Might Also Like

• Openais own Test Models Escaped Their Sandbox and Breached Hugging Face to Cheat on a Benchmark
• Handbook md Benchmark Surge ai the top Agent Still Fails 64 of Company Policy Tasks
• Hugging Face Speech to Speech Open Source Local Voice Agents is the Openai Realtime Clone you can run on Your own gpu
• Ggml Llama cpp Joins Hugging Face and Honestly it was Only a Matter of Time
• Pollen Robotics Reachy Mini a 299 Desktop Humanoid That Runs 1 7m Hugging Face Models