By Keshavan Thiagaraja
Two days after Google announced that its strongest model would be available only to a guarded circle of cyber defenders, Mistral AI took the opposite path. On 29 October the Paris laboratory released the weights of Mistral Large 4, its flagship trillion-parameter model, for anyone to download, inspect and run on their own machines.
The numbers explain why this matters. After three more weeks of reinforcement learning since its public preview, Mistral Large 4 now leads Cognition’s FrontierCode 2.0 benchmark with a score of 19.7 percent, and Terminal-Bench 4.0 at 41 percent, thirteen points ahead of its own preview while using thirty percent fewer output tokens. On the Artificial Analysis Coding Agent Index it matches the leading open-weight model, GLM-5.3, at 48 percent while consuming 47 percent fewer tokens per task. In Surge Code’s blind human evaluations, graders preferred it over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, ranking it behind only Claude Opus 5.5.
The security story is equally pointed. Mistral Large 4 scores 51 percent on the Artificial Analysis Cyber Index, which tests whether a model can find vulnerabilities in a codebase, reproduce them, and patch them without breaking anything. It handles these defensive tasks without unnecessary refusals, while refusing 96.6 percent of malicious cyber requests, the highest rate Mistral has measured among open models. Booz Allen Hamilton, Cisco, Cloudflare, Fortinet, Rubrik and Zscaler all lent their names to the release.
This is the same capability class that Google decided to ration. One laboratory concluded that a model this good at finding and fixing flaws must be kept behind vetting and contracts. The other concluded that defenders are better served when everyone can hold the same weapon. Both arguments are serious, and the industry now gets to watch them tested side by side in public.
The enterprise numbers round out the case. Mistral Large 4 reaches 73.3 percent on the Finch finance-workflow benchmark, 84.5 percent on BrowseComp for web research, and resists 97.1 percent of attacks on the B3 Agent Security benchmark, the best result among open-weight models. Three reasoning modes, low, medium and high, let teams trade cost against precision per task.
The sovereignty argument is the deeper play. The model was trained from scratch on European infrastructure with a clear lineage, and Mistral is explicit about the audience: governments, banks and regulated industries that cannot send their data to someone else’s cloud. La Banque Postale, a public bank, says technological sovereignty is not optional for maintaining customer trust. A vetted cybersecurity variant is also in preview with partners such as Orange Cyberdefense under contractual safeguards.
For leaders, the practical question has changed. It is no longer whether an open model is good enough for serious work, because this one clearly is. The question is what your organisation does with a frontier model it can fully control: audit it, tailor it with proprietary data, and run it where your rules apply, or keep renting capability on someone else’s terms.
Openness and caution are now both on the board. The organisations that learn to evaluate, govern and deploy open-weight models today will be the ones setting their own terms tomorrow, whichever philosophy wins.
Think mathematically. Engineer intelligently. Lead strategically. Build responsibly. Scale globally. Create real business value through artificial intelligence.
#ArtificialIntelligence #AIStrategy #MachineLearning #Research #ITImpress
