Mistral releases open weights for its upgraded flagship model

Written by

Picture of فريقنا

فريقنا

Communications Consultant

French AI developer Mistral has released open, downloadable weights for an improved version of its flagship model, Mistral Large 4, also known as Le Chonk. The multimodal model features a total of one trillion parameters with approximately 52 billion active parameters, focusing on coding, reasoning, and defensive cybersecurity, alongside a specialized partnership with Orange Cyberdefense.

The open weights of Mistral Large 4 provide a robust option for European enterprises looking to run AI models on their own infrastructure for coding and defensive cybersecurity tasks.

Release of downloadable weights and mixture-of-experts architectural specs
Coding benchmark results and cybersecurity indicators
Collaboration with Orange Cyberdefense, licensing terms, and pricing
Importance of the model for European enterprises and independent verification requirements
Frequently asked questions

Release of downloadable weights and mixture-of-experts architectural specs

French AI development company Mistral released the open weights for an upgraded version of its flagship model, “Mistral Large 4″—also known as “Le Chonk”—on October 29, 2026, according to company statements and Reuters reports. The model was first introduced in a public preview on October 6, but the new development on October 29 marks the availability of direct downloadable weights accompanied by additional training and evaluations, rather than the initial announcement.

Mistral describes the system as a multimodal model based on a “mixture-of-experts” architecture featuring one trillion total parameters, with roughly 52 billion active parameters during processing. The model focuses primarily on computer programming tasks, logical reasoning, and defensive cybersecurity, while enabling organizations to run the model directly on their own infrastructure.

From a legal and technical standpoint, it is essential to note that the term “open weights” does not necessarily mean the original training dataset, the weight usage license, or the commercial use terms are entirely free of contractual restrictions.

Coding benchmark results and cybersecurity indicators

Mistral reported that the upgraded version achieved a score of 41% on the Terminal Bench 4.0 benchmark, representing a 13 percentage point increase compared to the previous preview version, and scored 48% on the Artificial Analysis coding agent index while consuming fewer tokens than a competing model.

In the field of digital security, the company announced a score of 51% on the Artificial Analysis cybersecurity index, alongside a 96.6% refusal rate for selected prompts related to malicious cyber activities. Although some of these benchmark guides come from external evaluation entities, the choice of comparison models and usage conditions were presented within Mistral’s own announcement.

These benchmark scores do not mean the system successfully executes every programming or security task flawlessly in real-world work environments, nor do high test refusal rates prove that the model’s misuse or circumvention of safeguards is impossible.

Collaboration with Orange Cyberdefense, licensing terms, and pricing

The French company also revealed close cooperation with partners in the cybersecurity sector; Orange Cyberdefense is gaining trial access to a specialized version of the model under controlled conditions for use in defensive workflows such as authorized vulnerability assessments. Mistral emphasized that this access is subject to prior auditing and strict contractual constraints, representing a distinct commercial application integrated into the new Large 4 launch.

The public weights are released under the “Mistral Open Weights Ecosystem License”, a modified MIT-style agreement that requires a separate agreement for commercial cloud model providers. Hosted API pricing is set at $0.70 per million input tokens and $2.10 per million output tokens, with discounts for cached inputs.

Importance of the model for European enterprises and independent verification requirements

This launch is particularly significant because European organizations seeking self-hosted AI models now have a publicly available option capable of handling complex technical work. This may help meet data localization and vendor control requirements, although running and maintaining a model of this massive scale still requires specialized computing resources and advanced security engineering.

The next steps involve independently reproducing the coding and cybersecurity evaluations, assessing practical licensing terms, and gathering field evidence from authorized enterprise deployments.

When analyzing this development within the open-weight foundation model, cybersecurity, and developer tool sectors in France, a key distinction emerges between the initial preview announcement on October 6 and the October 29 event, which marked the actual downloadable weights availability of the “Mistral Large 4” model, nicknamed “Le Chonk”, following extensive additional training and evaluations.

Although the mixture-of-experts architecture featuring one trillion total parameters and roughly 52 billion active parameters gives European enterprises an advanced technical option for self-hosting and data sovereignty protection, running a model of this scale requires elite computing infrastructure and specialized engineering and security expertise. Furthermore, standard benchmark results such as Terminal Bench 4.0, Artificial Analysis indices, and the 96.6% malicious prompt refusal rate still require independent verification in real production environments, especially as Orange Cyberdefense begins testing the specialized version.

Frequently asked questions

Question: What did Mistral announce regarding Mistral Large 4 on October 29?
Answer: The company released the open, downloadable weights for an upgraded version of the model nicknamed Le Chonk following additional training, building upon its initial preview release on October 6.

Question: How many parameters does Mistral Large 4 have?
Answer: The multimodal model relies on a mixture-of-experts architecture featuring one trillion total parameters with approximately 52 billion active parameters during processing.

Question: What is the nature of the partnership between Mistral and Orange Cyberdefense?
Answer: Orange Cyberdefense receives controlled and audited trial access to a specialized version of the model for use in defensive cybersecurity tasks and authorized vulnerability assessments.

Question: What are the open-weights licensing terms for Mistral Large 4?
Answer: The weights are subject to the Mistral Open Weights Ecosystem License, a modified license that requires a separate commercial agreement for companies offering the model as a commercial cloud service.

شارك هذا الموضوع:

شارك هذا الموضوع:

اترك رد

الفئات

المنشورات الأخيرة