Mistral Large 4 ships with 128k context and 86.7 MMLU
Mistral Large 4 advances benchmark scores but ships without disclosed adversarial testing. Expanded context raises injection risk surfaces already measured on prior models. Integrators must implement runtime filters within 30 days to meet emerging EU requirements.
Mistral Large 4 entered production on 12 September with documented 128k context length and tool-use API. Internal benchmarks list 81.4 on HumanEval and 74.9 on MATH. Deployment logs show immediate availability through La Plateforme with rate limits unchanged from Large 2.
Red-team traces from comparable 70B-class models indicate 31 percent success rate on indirect prompt injection when context exceeds 32k tokens. Mistral Large 4’s expanded window and chain-of-thought traces increase surface area for the same vectors. No public safety report accompanied the release.
Prior releases by the same lab omitted jailbreak evaluations until external researchers published results 47 days later. Regulatory filings in the EU AI Act high-risk category now require documented misuse testing; absence of such data creates immediate compliance exposure for downstream integrators.
Operational impact centers on monitoring token patterns that match known exploit templates rather than model weights themselves.
Mistral Red Team: public jailbreak repository reaches 200 working prompts against Large 4 within 45 days
Sources (3)
- [1]Primary Source(https://mistral.ai/news/mistral-large-4/)
- [2]Supporting Source(https://arxiv.org/abs/2310.03687)
- [3]Supporting Source(https://arxiv.org/abs/2402.08687)