Shieldstral 1.0
Shieldstral is a downloadable text and image moderation model. Instead of using only a fixed list of categories, it evaluates a policy question supplied in natural language. Its single-token yes/no decision can be converted into a continuous score.
Preview
Ask one clear policy question
Shieldstral is based on Ministral-3-3B-Base-2512 and a Pixtral vision encoder. Its card recommends one policy per question. Applications can use token probabilities to set their own decision threshold. Text-only, image-only and combined inputs use the same policy interface.Shieldstral 1.0 3B model card
Run the released weights
The official model page publishes weights but no Mistral serverless model identifier. It lists prompt checks, response checks, prompt-response pair classification and refusal detection. This directory therefore records weight access, without asserting a hosted moderation endpoint for Shieldstral.Shieldstral 1.0 model page
Evaluate policy wording and thresholds together
Practical guidance. Create examples near the boundary of each policy, including permitted material that resembles a violation. Changing the question can change the decision. Record the question, threshold and model version together so that a later policy revision can be evaluated against the same examples.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Preview |
| Launch date | |
| Access | Downloadable |
| Model identifiers | Shieldstral-1.0-3B |
| Context window | 32,768 tokens The model card recommends staying within its 32k training range; its example serves 32768 tokens. It does not recommend the theoretical 256k limit. Shieldstral 1.0 3B model card |
| Initial release stage | Public preview.Shieldstral 1.0 model page |
| Published size | 3.8B parameters in the model catalog; repository named 3BShieldstral 1.0 model pageShieldstral 1.0 3B model card |
| License | Apache 2.0Shieldstral 1.0 model page |
Related models
- Ministral 3 3B
Mistral AI · Ministral 3
The smallest Ministral 3 release pairs a 3.4B language model with a 0.4B vision encoder and offers multiple training variants.
Launch date not confirmedAPI · Downloadable - Mistral Moderation 2
Mistral AI · Mistral Moderation
A multilingual moderation model for long conversations, harmful-content detection and jailbreak checks.
Launch date API