Mistral Moderation 2
Mistral AI
Mistral AI's text moderation model, served through a classifier endpoint rather than as a chat model. It scores text against the policy categories Sexual, Hate and Discrimination, Violence and Threats, Dangerous, Criminal, Self-Harm, Health, Financial, Law, PII and Jailbreaking, returning per-category scores that the caller thresholds itself. The service exposes one endpoint for raw text and one for conversational content, supports batching, and backs Mistral's Custom Guardrails config (moderation_llm_v2). It replaces mistral-moderation-2411, which was deprecated on March 31, 2026.
Strengths
- Classifies text across eleven named policy categories, including a dedicated Jailbreaking category
- Returns raw per-category scores so callers can set their own thresholds instead of a fixed verdict
- Two endpoints: one for raw text chunks and one for conversational content
- Batched moderation supported - the raw-text endpoint accepts a list of strings, and /v1/batch is listed as a feature
- Backs Mistral's Custom Guardrails, which apply input moderation declaratively in the request and return 403 when triggered, with per-category threshold overrides
- Listed at Price: Free on the vendor's model page
- Vendor describes strong performance on complex multilingual data including long multi-turn conversations
Caveats
- Mistral publishes no benchmark scores for this model - there is no evaluation table on the model page or the guardrailing docs, so accuracy claims are unverifiable from vendor text
- Context is advertised only as '128k'; the vendor never states the exact token count, so the real limit could be 128,000 or 131,072
- Text only - the moderation endpoints classify text chunks and conversational content, with no image or audio moderation documented
- Closed weights with no published licence; it can only be used through Mistral's API, not self-hosted
- Mistral warns that it intends to keep changing the underlying model, and that 'Custom policies that depend on category_scores can require recalibration'
- Guardrails apply input moderation only - they run before the request reaches the model and do not screen model output
- Vendor
- Mistral AI
- License
- Proprietary
- Release Date
- 2026-03-01
- Modalities
- text
Capabilities
Resources
Pricing
The vendor's model page lists Price: Free. It is a Premier-tier model available through the Mistral API; the Moderations endpoint is /v1/moderations and batch requests go through /v1/batch.