Mistral AI released the weights of Large 4 on Hugging Face on 29 October, about three weeks after previewing the model. The trillion-parameter multimodal model, with 52 billion active parameters, is claimed to score 79% on CyberGym-E2E, which tests finding, reproducing and patching memory-safety bugs, against 76% for DeepSeek V4.1 Flash. The figures are company-reported, and early government and enterprise users are reported to be running automated penetration testing with it.

[S1] Mistral releases open weights for its 1 trillion-parameter Large 4 model · [S2] Mistral Large 4: open weights

From developments to evidence and analysis

Background

No explicit relationship is recorded.

Evidence

No explicit relationship is recorded.

Follow-ups / later versions

No explicit relationship is recorded.

Related questions

Primary AI safety & evaluationSecondary Cyber & critical infrastructure

Cite this article

AI & Security Observatory（2026-10-29）「Mistral、1兆パラメータのLarge 4の重みを公開　サイバー防御評価CyberGym-E2Eで79%と主張」『AI & Security Observatory』、第1.0版、最終更新2026-10-31。https://ai-security-observatory.org/ja/brief/brief-20261031-10/

AI & Security Observatory (2026-10-29). Mistral releases open weights for its trillion-parameter Large 4, claiming 79% on the CyberGym-E2E cyber-defence benchmark . AI & Security Observatory, version 1.0. Updated 2026-10-31. https://ai-security-observatory.org/en/brief/brief-20261031-10/

BibTeX · JA · RIS · JA · BibTeX · EN · RIS · EN

Version 1.0 · Publication date 29 Oct 2026 · Updated 31 Oct 2026

Revision history
  1. Earlier revisions have not been recorded.