Mistral’s Large 4 release lands amid a broader AI spending boom that has investors cheering fresh financing while questioning who will ultimately bear its cost. On Tuesday, CNBC’s Jim Cramer warned that the data-center rally did not erase the sector’s debt problem: “The rates are getting so high that the return is getting uncertain.”¹

Earlier this month, the Paris-based lab offered a limited preview of its trillion-parameter Large 4, nicknamed “Le Chonk.” After further reinforcement-learning work, Mistral said its Terminal-Bench 4.0 score rose from 28% to 41%, drawing level with China’s GLM-5.3 on that coding-agent test.²

On Thursday, Mistral released the model’s weights publicly through Hugging Face and its API, allowing companies and governments to run it in their own systems rather than depend on a remote provider. The company framed that access as a cyber-defense issue, saying ML4 gives institutions the ability to “download, deploy, and own frontier AI” for threats across code and IT infrastructure.³

That is the core of Mistral’s argument: control cannot be switched off. Pierre Stock, its vice president of science, said an open-weight model gives organizations access to defensive intelligence “that can’t be cut off by definition.”³ Mistral also says Large 4 uses 47% fewer tokens per task than GLM-5.3, an efficiency claim aimed at customers in regulated industries seeking both lower costs and continuity.²

The release nonetheless exposes the limits of the victory lap. Mistral says Large 4 is the best open-weight model for coding and cyber work, and that it is more likely to reject malicious cyber requests than rival open models.⁴ But closed systems remain ahead: Anthropic’s Mythos 5.1 scored 60.9% on the same benchmark, versus Large 4’s 41%.²

Mistral’s gamble, then, is not that it has already beaten America’s closed-model giants. It is that sovereignty, deployability and cyber resilience will become valuable enough that Europe’s open challenger no longer needs to.