Mistral opens Large 4, its giant model for coding and security
Mistral AI releases weights of Large 4, its 1 trillion parameter multimodal model. The company primarily targets code, cybersecurity, and broader business uses, with enhanced deployment control.
Mistral AI has published the weights of Large 4, its 1 trillion parameter multimodal model, on Hugging Face, while making it accessible via its API and in Vibe Code. Announced on October 6, this model called “Le Chonk” mainly targets agentic software development and cybersecurity, two areas where cost and data control matter as much as raw performance. This release marks another step in the company’s strategy to combine open models, professional use, and deployment in controlled environments.
Un modèle plus ouvert, mais pas plus léger
Large 4 has 1 trillion parameters, including 52 billion active ones. This architecture maintains high analytical power while activating only part of the model per request, directly impacting token consumption. Regarding agentic programming, Mistral claims Large 4 rivals GLM-5.3 while using on average 47% fewer tokens per task. This detail matters for companies, as AI agent costs can quickly escalate unnoticed.
The model also benefited from three additional weeks of reinforcement training after its October 6 presentation, incorporating feedback from clients and partners. Mistral says it added three reasoning levels — low, medium, and high — to adjust model behavior according to use cases. This approach targets daily use rather than a one-off showcase, ending up in internal tools, code assistants, or business processing chains.
Les performances comptent, mais la facture aussi
On Terminal-Bench 4.0, a benchmark evaluating agents capable of executing programming operations in a terminal, Large 4’s score reportedly rose from 28% to 41%, according to Mistral. The company thus claims to reach GLM-5.3’s level on this indicator. Artificial Analysis ranks Large 4 as the top open model for agentic software development, and Mistral presents it as the best open model developed in the West.
The company also highlights training the model on 3,800 NVIDIA Grace Blackwell GPUs in its European data centers. This is significant: it gives an idea of the industrial scale needed to build a model of this size without necessarily having the budgets of the largest market players. Arthur Mensch, co-founder and CEO of Mistral, recalls that the company raised about $6 billion in three years. Enough to fund serious ambition without handing out freebies at every benchmark.
Pourquoi la cybersécurité est au cœur du positionnement
Mistral also targets cybersecurity, with announced uses for code analysis, threat detection, and IT infrastructure review. The model scores 51% on the Artificial Analysis Cyber Index, which notably measures its ability to identify, reproduce, and fix vulnerabilities. The company also reports an average 96.6% rejection rate against malicious cyber requests, according to several specialized test sets.
Publishing the weights changes the game for organizations wanting to run AI in their own environment. Companies and administrations can thus avoid exclusive dependence on a remote service, which is especially interesting for regulated sectors. Mistral presents this option as a way to better control AI uses and data flow. In practice, this can make the difference between a simple online tool and a system integrated at the heart of a sensitive architecture.
Des usages qui dépassent largement le code
The model is not limited to developers. In industry, Mistral cites transforming 2D plans into 3D models usable in computer-aided design (CAD) software. In finance, Large 4 is presented as capable of handling complex spreadsheets, analyzing large data volumes, and exploiting financial publications. The model’s multimodal capabilities can also analyze satellite and aerial images to detect fires, monitor crop conditions, or spot risks on electrical infrastructures.
The weights are published under the Mistral Ecosystem Open-Weight License, a license derived from MIT that governs redistribution and commercial use of the model. This is a strategic choice: opening access while keeping a precise legal framework. For professional users, the promise is clear: more flexibility, but not total freedom without safeguards.
By the numbers
- 1 trillion parameters for Large 4, according to Mistral.
- 52 billion active parameters in the model’s architecture.
- 47% fewer tokens per task on average in agentic programming.
- 3,800 NVIDIA Grace Blackwell GPUs used for training in Europe.
- $6 billion raised by Mistral in three years, according to Arthur Mensch.