Mistral Releases Large 4 “Le Chonk,” a 1-Trillion-Parameter Open-Weight AI Model

French artificial intelligence company Mistral AI has released the weights for Mistral Large 4, its flagship one-trillion-parameter AI model, giving developers and businesses the option to download, customize, and deploy the technology on their own infrastructure.

The model, internally nicknamed “Le Chonk,” was first introduced in public preview on October 6, 2026. Mistral released an improved version with downloadable weights on October 29, following several weeks of additional training and feedback from developers and early-access partners.

The release marks another step in Mistral's effort to compete with leading AI developers while offering an alternative to systems that can only be accessed through a company's hosted services.

What is Mistral Large 4?

Mistral Large 4 is a multimodal, open-weight artificial intelligence model designed for a range of demanding tasks, including software development, reasoning, cybersecurity, and enterprise workflows.

The model contains approximately one trillion parameters, with around 49 billion active parameters per token in its mixture-of-experts architecture. This design allows the system to draw on a large pool of model capacity without activating every parameter for each operation.

It also supports text and image inputs and is designed to work across more than 160 languages.

Mistral says the model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its European data centers. The company has positioned the system as part of its broader effort to strengthen European control over advanced AI infrastructure and technology.

Mistral Large 4 is now available as an open-weight model

One of the most important aspects of the release is the availability of the model's weights.

Unlike a traditional closed AI service, an open-weight model allows developers to download its trained parameters and operate the system in environments they control. This gives organizations more flexibility over deployment, customization, infrastructure, and data handling.

Mistral Large 4 can be accessed through Mistral Studio and its API, as well as platforms such as OpenRouter. Developers can also download the weights through Hugging Face and deploy the model using tools including vLLM and SGLang.

The model is released under Mistral's Ecosystem Open-Weight license, a modified MIT license. It permits use, modification, fine-tuning, and commercialization, although businesses operating a Model-as-a-Service offering must obtain a separate commercial license from Mistral.

This distinction matters for companies considering whether to build AI products around the model or offer hosted access to it.

A stronger focus on AI-powered software development

Software engineering is one of the main areas Mistral is targeting with Large 4.

The company says additional reinforcement learning after the initial preview improved the model's performance on coding and agent-based development tasks.

In its October 29 announcement, Mistral reported a 48% score on the Artificial Analysis Coding Agent Index, describing the result as effectively tied with the leading open-weight model while using 47% fewer tokens per task on average.

The company also reported a 41% result on Terminal-Bench 4.0, a 13-percentage-point improvement over the public preview.

These results are relevant to developers building AI agents that can understand codebases, work through terminal environments, identify problems, and make software changes.

Mistral also demonstrated an agentic web-development workflow in which Large 4 built an interactive 3D visualization of roughly 19,000 satellites and pieces of orbital debris. The demonstration showed the model retrieving data, writing an application, testing it in a browser, and making changes based on identified issues.

While demonstrations do not establish how reliably a model will perform across every production environment, they illustrate the direction Mistral wants to take with AI-assisted software development.

Cybersecurity is another major focus

Mistral is also positioning Large 4 as a model for defensive cybersecurity work.

The company says the model can help security teams investigate vulnerabilities, reproduce security flaws, validate potential exploits in controlled environments, and develop patches.

According to Mistral's October 29 announcement, Large 4 scored 51% on the Artificial Analysis Cyber Index. The benchmark evaluates capabilities related to finding vulnerabilities in software, reproducing them, and patching them without breaking existing functionality.

Mistral also reported an average refusal rate of 96.6% on selected malicious cybersecurity requests, based on evaluations involving JailbreakBench, StrongREJECT, and AgentHarm.

These figures are company-reported results and should be understood in the context of the benchmarks used. They do not mean the model is immune to misuse or that it can replace experienced security professionals.

The company says it has worked with cybersecurity organizations and enterprise partners to evaluate the model in practical environments. Organizations mentioned in its announcement include Booz Allen Hamilton, Cisco, Cloudflare, CrowdStrike, Elastic, Fortinet, and Orange Cyberdefense.

Mistral also acknowledged that its monitoring of real-world usage identified potential opportunities for misuse. It said it was adjusting its safeguards while providing specialized cybersecurity variants to vetted enterprise partners under contractual restrictions.

This is an important part of the open-weight AI discussion. Giving organizations greater control over an AI model can make it easier to adapt the technology for legitimate security work, but it also makes responsible deployment and access controls especially important.

What open-weight AI means for businesses

The release of Mistral Large 4 reflects a growing interest in AI systems that organizations can operate with greater independence.

Businesses using closed AI platforms often depend on external providers for model access, infrastructure, pricing, and service availability. Open-weight models provide another option, allowing organizations to evaluate whether they can host and customize AI systems within their own environments.

For companies handling sensitive information, this can be particularly valuable. Internal documents, proprietary code, customer records, and security data may be subject to strict privacy, compliance, or data-residency requirements.

Self-hosting does not automatically guarantee security or regulatory compliance. Organizations still need appropriate infrastructure, access controls, monitoring, and governance. However, having the ability to control where a model runs can provide more flexibility when designing those safeguards.

Mistral Large 4 is also available through Microsoft Foundry, giving enterprise customers another deployment route, including private deployment options for Azure developers.

Pricing and availability

Mistral has made Large 4 available through its API and several developer platforms. Organizations that want more control can download the model weights and deploy them on compatible infrastructure.

At the time of its October 29 release announcement, Mistral listed API pricing at:

1. Input: $0.70 per million tokens
2. Cached input: $0.07 per million tokens
3. Output: $2.10 per million tokens

Actual deployment costs will depend on the infrastructure, hardware, workload, and configuration an organization chooses.

The model's size also means that self-hosting is not automatically a simple or inexpensive option for every developer. Organizations should evaluate hardware requirements, inference performance, operational costs, and security responsibilities before choosing a deployment approach.

The growing competition in open-weight AI

Mistral Large 4 arrives as competition intensifies among companies developing advanced AI models.

While some major AI providers focus on closed systems delivered through hosted platforms, Mistral is making a case for models that customers can inspect, adapt, and run under their own operational policies.

The company's European infrastructure and emphasis on AI sovereignty are also part of that strategy. For governments, financial institutions, and other organizations concerned about technology dependence, control over AI deployment can be as important as model performance.

However, benchmark results alone will not determine whether Large 4 becomes a major enterprise platform. Reliability, total operating cost, ecosystem support, security, and performance in real-world applications will all influence adoption.

For now, the release gives developers and businesses another significant open-weight model to evaluate, particularly in software engineering, AI agents, and defensive cybersecurity.

As organizations look beyond simply using AI chatbots, models that can be customized and integrated into internal systems may become an increasingly important part of business technology infrastructure.