Glossary · Term

Ministral

← all terms

Definition

Plain language

A smaller model in Mistral's family of open-weight language models.

As stated in the literature

Mistral AI's line of compact open-weight models, distinct from the larger dense Mistral and Mixtral variants; used as one of several backbones in reward-hacking and tool-deference experiments.

Why it matters: Its compact size makes it a practical choice for testing ideas cheaply and checking whether findings hold on smaller models, not just big ones.

For example, a study might run its experiment on Ministral to see how a smaller open model behaves, then compare against larger models in the same family.

Heard on the show

“" And it reproduces across model families — Qwen, Mistral, Ministral, a Llama variant.”
Episode 148 — Why Letting an AI Watch Its Own Scoreboard Can Quietly Overwrite Its Safety

Mentioned in 1 episode

  1. 148
    Why Letting an AI Watch Its Own Scoreboard Can Quietly Overwrite Its Safety

Related terms