Mistral opened its trillion-parameter model; Google kept its own closed

Just two days after Google announced its most capable model would be available only to a protected circle of selected internet security defenders, Mistral AI chose the opposite path. On October 29, the Paris-based laboratory released the weights of its flagship trillion-parameter model, Mistral Large 4, so anyone can download, study and run it.

After three extra weeks of reinforcement learning, the model leads Cognition's FrontierCode 2.0 benchmark at 19.7 per cent and reaches 41 per cent on Terminal-Bench 4.0, thirteen points ahead of its predecessor while using thirty per cent fewer output tokens. It ties the leading open-weight model GLM-5.3 at 48 per cent on the Artificial Analysis Coding Agent Index, and in Surge Code's blind human evaluations raters placed only Claude Opus 5.5 ahead.

On security it scored 51 per cent on the Artificial Analysis Cyber Index, which measures whether a model can find and patch vulnerabilities without breaking anything, while rejecting 96.6 per cent of malicious internet requests, the highest rate measured among open models. Booz Allen Hamilton, Cisco, Cloudflare, Fortinet, Rubrik and Zscaler lent their names to the release.

This is the same class of capability Google decided to withhold. One laboratory concluded such a model should stay behind research contracts; the other that defenders are better served when everyone can carry the same weapon. Enterprise figures complete the case: 73.3 per cent on the Finch finance-workflow benchmark, 84.5 per cent on BrowseComp, and 97.1 per cent of attacks resisted on the B3 Agent Security benchmark.

The sovereignty argument is the deeper move. Trained from the ground up on European infrastructure, the model is aimed openly at governments, banks and regulated industries that cannot send data to someone else's cloud. Public-sector bank La Banque Postale says technological sovereignty is not optional for preserving trust.

For leaders the practical question has changed: whether an open model is sufficient for serious work is settled. What matters now is what your organisation does with a frontier model it can control: audit it, customise it, run it where its rules apply, or keep renting on someone else's terms.