Mistral Unveils Large 4 “Le Chonk,” a Trillion-Parameter AI Model

Key Notes

• Mistral Large 4 is available in public API preview with a one-million-token context window.
• The model targets coding, cybersecurity and other enterprise workflows, with performance claims specific to published evaluations.
• Open weights are planned for late October, while the license remains unpublished.

Mistral has introduced Large 4, a trillion-parameter AI model nicknamed “Le Chonk,” with a public API preview available from October 6. The French company is positioning its new flagship around coding, cybersecurity and professional work, while promising an open-weight release by the end of the month.

The launch gives European businesses another model to evaluate for demanding tasks and, eventually, deployment under their own control. It also comes with a distinction that matters immediately: developers can test the hosted preview now, but the downloadable weights and their license have not yet been published.

What Is Mistral Large 4?

Mistral’s model documentation describes a multimodal mixture-of-experts system with 1.05 trillion total parameters, 49 billion active parameters and a 1.6-billion-parameter vision encoder. It lists a one-million-token context window and the API identifier mistral-large-4.

The active-parameter figure describes the portion of the model used during computation, rather than a smaller standalone model. A mixture-of-experts design routes work through selected components instead of activating the entire network for every token. The trillion-parameter headline therefore says more about overall scale than the resources required for each individual calculation.

That architecture does not make self-hosting requirements trivial. The documentation currently leaves its GPU-memory estimate unspecified. Organizations considering a private deployment will need the released weights, supported inference software and hardware guidance before they can judge whether running it themselves makes operational sense.

What Can Developers Use Today?

The preview is accessible through Mistral Studio and its API. The model card lists structured outputs, function calling, document question answering, batching and support for agents and built-in tools. These are the components developers need to connect a model to software workflows rather than use it only as a conversational assistant.

A large context window can accommodate substantial documents or project material, but fitting information into a request does not guarantee that every relevant detail will be retrieved correctly. For a document-heavy application, the meaningful test is whether the system cites the right evidence, follows instructions and handles contradictions consistently.

Mistral’s announcement says the model is still improving through ongoing training. Preview results should therefore be treated as a snapshot. Teams evaluating it for production should record the model version and configuration so they can identify whether later changes improve or disrupt their workflows.

How Much Does Le Chonk Cost?

At publication, the model card displays discounted preview rates of $0.68 per million input tokens, $0.07 for cached input and $2.09 for output. Alongside those figures, it shows undiscounted rates of $1.36, $0.14 and $4.18 respectively.

These are API token prices, not a subscription allowance or an estimate of self-hosting costs. A workflow’s final expense also depends on how many steps it takes, how much context it repeats and how often an answer needs correction. A low output-token rate can lose its advantage if a task requires repeated unsuccessful attempts.

Where Mistral Claims a Performance Advantage

Mistral reports 61.7% on DeepSWE 1.1, 28.3% on Terminal-Bench 4 and 59.9% on AutomationBench. Its announcement also cites an 82% result on a cybersecurity task involving reproducing and patching a real vulnerability, plus 93% on Cybench’s security challenges.

The company’s case is strongest around particular enterprise tasks, rather than a claim of winning every benchmark. Its published comparisons cover software engineering, financial and legal work, scientific coding and visual grounding. Mistral says its visual-grounding results can exceed leading closed models in selected evaluations.

Those results need to be read with their testing conditions. Coding outcomes depend on the agent setup, available tools, reasoning budget and evaluation rules. A result on vulnerability reproduction can also be affected by whether a competing model refuses the task. That difference matters to a security team, but it should not be mistaken for a universal ranking of intelligence.

Cybersecurity Access Is Part of the Strategy

Mistral says vetted cybersecurity partners and state authorities are testing the same model with reduced moderation and expanded cyber capabilities before the weights release. The public preview and that controlled testing program are therefore different access arrangements.

The company argues that provider-level refusals can interfere with legitimate vulnerability research or incident response. Our earlier coverage of Mistral’s banking push examined the demand for a European alternative to restricted security models. Large 4 now puts that argument alongside a publicly testable general-purpose model.

Greater control also moves responsibility toward the organization operating the system. A security deployment needs clear authorization boundaries, reliable records of its actions and checks before changes reach live infrastructure. Access to capable software is only one part of building a dependable incident-response process.

Why European Infrastructure Matters

Mistral says it trained Large 4 from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own European data centers and serves the public preview from that infrastructure. The company is presenting the combination of model development, hosting and future self-deployment as an answer to concerns about dependency on outside providers.

For a business, sovereignty is a practical question about where data is processed, who controls service access and whether a critical system can keep running if a supplier changes its policies. European hosting can help address some of those requirements. It does not, by itself, prove that a model is more accurate or that every deployment meets a particular compliance obligation.

When Will the Open Weights Arrive?

Mistral’s stated target is the end of October. The model card currently marks both weights and license as “coming soon,” so it is premature to describe Large 4 as available for unrestricted download or commercial self-hosting under a specified license.

The distinction between open weights and a fully open-source release remains useful. Weights provide the learned model parameters; the eventual license determines permitted uses and conditions. The release package and documentation will also show what customers need to reproduce deployment outside Mistral’s hosted service.

The broader market includes specialized open models as well as large general-purpose systems. Cloudflare’s Clef models, for example, focus on structured decisions inside applications. Large 4 addresses a different problem: providing a broad model for multi-step professional work. Organizations may ultimately use both kinds of component in the same system.

For now, Le Chonk is an invitation to test Mistral’s renewed capability claims against real assignments. The hosted preview provides the first evidence customers can gather themselves. The planned weights release will determine how far that capability can travel beyond Mistral’s infrastructure, and on what terms.

Disclaimer: AIstify is an independent media brand owned and operated by NuvexMedia LLC, publishing news, research, and insights on artificial intelligence, emerging technologies, automation, and related industries. NuvexMedia LLC invests in and collaborates with companies across the AI, technology, software, and digital innovation sectors. These relationships do not influence AIstify’s editorial coverage, and the publication maintains full editorial independence to provide accurate, timely, and objective information. © 2026 NuvexMedia LLC. All rights reserved. This content is for informational purposes only and should not be considered legal, tax, investment, financial, or other professional advice.