Today

Mistral releases Large 4's weights and pitches them for cyber defense, and the same post says the preview showed signs the model could help real attacks

Concept illustration: at the top of a pale backdrop panel, a stepped block icon in yellow, orange, red and black sits beside the lettering "MISTRAL AI" above grey bars; on the left, a person wearing a lanyard badge holds a box marked with a beetle symbol in front of an open server rack, with an open cardboard box on the floor beside them. In the centre, a red laptop lettered "LARGE 4" sits on a table under a speech bubble holding grey bars, and next to it a second beetle-marked box rests at the top of a slanted board that runs from the table edge down to another server rack on the right.

Why it matters: Our inference: teams that self-host Large 4 keep the guardrails in the weights they downloaded. Mistral says the preview showed the model might help real attacks and that it is revising its guardrails, but those changes won't reach copies already downloaded.

French AI company Mistral AI released the weights of its flagship model, Large 4, on October 29.

This release had three more weeks of reinforcement learning than the API-only preview of October 6. Mistral reports 41% on Terminal-Bench 4.0, a test of terminal tasks, 13 percentage points above the preview, with 30% fewer output tokens. API pricing is US$0.70 per million input tokens and US$2.10 per million output tokens, about half the preview's US$1.36 and US$4.18 (Mistral, 2026-10-29; Mistral's preview announcement, 2026-10-06). The license allows free commercial use; only companies reselling the model as an API service need a separate license from Mistral.

Mistral pitches Large 4 for cyber defense, and the same post describes the risk. Mistral's argument is that defenders have to reproduce a vulnerability before they can fix it, and closed models often refuse that work. Large 4 refuses malicious security requests 96.6% of the time, the highest among the open models Mistral has measured, on prompts drawn from three public jailbreak and harm datasets. The same post says that while monitoring live traffic during the preview, "we see a potential for uplift to conduct real attacks", meaning the model could help people carry out real attacks, and that Mistral is revising its guardrails.

The two sources quote scores from different versions. Hugging Face, which hosts the weights, wrote in its same-day introduction that Large 4 scored 28.3% on the same terminal test and 93.3% on the agent-security test B3 (Hugging Face, 2026-10-29). Those two numbers match the ones in Mistral's preview announcement; the release scores are 41% and 97.1%.

Not proven yet: the scores and refusal rate are Mistral's own, not retested by anyone else How we checked →

Next to watch: Artificial Analysis publishing scores for the release version, and Mistral saying whether the change is API-side filtering or retrained weights. If Mistral re-releases modified weights, self-hosters who switch to them get the new guardrails. The inference then holds only for teams still running the old weights.

Source and checks

This item ran in the brief of Oct 30, 2026, alongside 13 other items. How we checked it, the quote checks and what was still missing that day are on that issue's How we checked page.

Cite this item

SecondSource (2026). "Mistral releases Large 4's weights and pitches them for cyber defense, and the same post says the preview showed signs the model could help real attacks." https://secondsource.io/en/news/2026-10-30-2/. Accessed access date.