Mistral Large 4 bets on open weights for enterprise

Mistral AI releases the weights of Mistral Large 4, its flagship model focused on enterprise and cybersecurity. The system can be downloaded, self-hosted, and aims for technical sovereignty, with a reduced usage cost.

Mistral AI published on Thursday the open weights of Mistral Large 4, its new flagship model designed for enterprise use and defensive cybersecurity. Named Le Chonk, this multimodal model arrives after three weeks of preview testing with corporate clients and public agencies. Its appeal is simple: instead of a service locked behind a cloud interface, organizations can download the weights, host it internally, and adapt it to their needs.

Des poids téléchargeables, pas une simple vitrine cloud

Weights are the parameters learned by the model during training; they determine how it responds and reasons. By making them available through platforms like Hugging Face, Mistral enables companies to self-host, thus running the system on their own private servers. Pierre Stock, Vice President of Science at Mistral, summarized the issue clearly: “I’m not comfortable living in a world in which open weights do not exist,” before mentioning dependency on providers and sudden service outages. Marjorie Janiewicz, Chief Revenue Officer, emphasized personalization, internal data governance, and business continuity.

This positioning appeals as much to CIOs as to security managers. In a context of tensions over digital sovereignty, the idea is not only to run a powerful model. It is also about keeping control over data, updates, and deployment pace. A model sleeping on your servers is less glamorous than a demo chatbot, but often more reassuring.

Un modèle très lourd, donc réservé aux grosses machines

On paper, Mistral Large 4 does not hold back. The system features 1.05 trillion parameters in total, 52 billion active per forward pass, and a context window of one million tokens. It was trained on about 3,800 NVIDIA Grace Blackwell GPUs, in more than 160 languages, and processes both text and images. This massive architecture enables agent use cases, i.e., long and multi-step workflows, but also requires a robust infrastructure to run locally.

Mistral highlights capabilities for processing complex documents, analyzing satellite or aerial images, and better grounding of responses to limit hallucinations. Evaluations cited by the company and Artificial Analysis also show good performance on SWE-bench and Terminal Bench, two benchmarks used to measure code auditing, pull requests, and threat detection. According to Artificial Analysis, the model scores 26 points on its Intelligence Index v5 in high reasoning mode, ahead of Meta Platforms Inc.’s Muse Spark Open, but still behind the leading Chinese open weights models DeepSeek V4.1 Flash and GLM-5.3.

Des coûts d’inférence revus nettement à la baisse

The launch is not only about capabilities. Artificial Analysis also mentions a 63% reduction in operating costs per benchmark compared to the preview version, dropping from $5.54 to $2.06 per indexed task. Mistral has simultaneously lowered its API prices to $0.70 per million input tokens and $2.10 per million output tokens. However, running such a large model locally will still require significant hardware investments.

Legally, the release is based on the Mistral Ecosystem Open-Weight License, an adaptation of the standard MIT open source license. It authorizes execution, modification, and redistribution of weights without arbitrary revenue caps but imposes a targeted restriction: actors who sell direct access to the model or post-training capabilities to third parties must negotiate a separate commercial agreement. Internal deployments and embedded enterprise applications remain exempt. The framework thus seeks to open usage without losing the entire business model.

Key points

  • Mistral Large 4 was released Thursday as open weights.
  • The model has 1.05 trillion parameters in total.
  • Artificial Analysis gives it 26 points on its Index v5.
  • Benchmark costs drop by 63% according to Artificial Analysis.
  • Self-hosting remains intended for companies capable of heavy infrastructure.
  • The Mistral license targets internal uses, not platform clones.

By the numbers

  • 1.05 trillion parameters — total model size, announced by Mistral.
  • 52 billion active parameters — per forward pass, according to Mistral.
  • 1 million tokens — context window, according to Mistral.
  • 3,800 NVIDIA Grace Blackwell GPUs — compute volume used for training.
  • 63% — benchmark cost reduction, according to Artificial Analysis.

To read