Mistral Large 4 (Le Chonk) is now Open Weight

Mistral released Mistral Large 4 (Le Chonk), a 1-trillion-parameter multimodal model, now accessible alongside improved coding, cybersecurity, and reasoning efficiency. It leads open-weight benchmarks in coding and protective safety during offering three reasoning modes for cost-performance balance. Fully open-weight and European-trained, it empowers sovereign AI alongside customizable, self-hosted deployment options.

We publish the weights of Le Chonk, advancing the frontier of open models in agentic coding and cybersecurity.

Earlier this month, we released Mistral Large 4 (Le Chonk), our flagship 1-trillion-parameter, natively multimodal model, in community preview. The reply to the preview has exceeded our expectations, and we're grateful to our developer community, early-access customers, and partners for their support, collaboration, and feedback.

Since then, we've put that feedback to activity during continuing to measure reinforcement learning, driving enormous improvements in coding and protective cybersecurity. Gains continue to accelerate, alongside no signs of saturation.

Today, we're releasing the weights of this improved Mistral Large 4. The new example additionally features significant gains in reasoning effectiveness that provision developers greater achievement during lowering conclusion costs.

Mistral Large 4 ranks as the best open-weight example for coding and among the world's most capable open models overall. Trained from scratch on European infrastructure alongside apparent lineage from inception to release, it gives organizations a trusted basis for Sovereign AI

Mistral is defining the sovereign AI category and foremost the open frontier. We're committed to construction the world's most capable models and releasing them openly, giving developers, enterprises, and institutions the power to own, customize, and authority their AI.

Built for the Builders.

Mistral Large 4 entered community preview alongside powerful results throughout foremost coding benchmarks. Just three weeks of additional reinforcement learning have since propelled it to the forefront of open-weight coding, taking the guide on FrontierCode and Terminal-Bench during matching the finest open-weight models on the Artificial Analysis Coding Agent Index alongside substantially greater token efficiency.

On Cognition's personal FrontierCode 2.0 benchmark, which assesses end-to-end code norm and whether a repository maintainer would merge the agent's change, Mistral Large 4 leads alongside a mark of 19.7%.

Pushing the pareto frontier for agentic coding.

Mistral Large 4 delivers frontier coding achievement alongside considerably small tokens, making advanced coding agents additional cost-effective to run at scale.

On the Artificial Analysis Coding Agent Index, Mistral Large 4 scores 48%, efficiently tied alongside the foremost open-weight model, GLM-5.3, during using 47% small tokens per project on average.

The gains broaden to terminal-based coding workflows. On Terminal-Bench 4.0, Mistral Large 4 improved by 13 percent points to 41% compared to its community preview, taking the guide among open-weight models during reducing output tokens by 30%.

To provision developers greater authority complete this tradeoff between achievement and cost, Mistral Large 4 now supports three reasoning modes:

Beyond benchmarks, Mistral Large 4's coding capabilities are validated by individual preference. In Surge Code's blind pairwise evaluations, Mistral Large 4 was preferred complete foremost open-weight models including GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash, ranking rearward lone Claude Opus 5.5 among the models evaluated.

Built for the Defenders. 

Since its preview, we've been operating alongside foremost cybersecurity vendors to measure Mistral Large 4 in real-world environments to test Mistral Large 4’s strengths in exposure reproduction and patching.

Mistral Large 4 scores 51% on the Artificial Analysis Cyber Index, which measures a model’s capability to detect vulnerabilities in a codebase, reproduce and validate them, and place them without breaking existing functionality. Crucially, Mistral Large 4 handles these lawful protective safety tasks without unnecessary refusals.

Safety has improved alongside capability. Our model’s built-in guardrails distinguishes between protective and malicious, insulting cybersecurity tasks. It retains powerful achievement on Cyber benchmarks, during achieving an average refusal charge of 96.6% on malicious cyber requests, the highest we measure among open models.

Here’s what earth leaders in cybersecurity have to say concerning Mistral Large 4:

In the period since the preview, we carefully monitored cybersecurity-related use in our live traffic. Though we observed the example refusing to comply alongside black-box attacks against live targets, we see a possible for uplift to behavior genuine attacks. We are modifying our guardrails accordingly, and volition continue updating them in the months to come. 

We are additionally providing endeavor scheme partners specified as Orange Cyberdefense, with preview admission to a specialized type of Mistral Large 4 optimized for advanced cybersecurity workflows, including intrusion testing, exposure assessment, and red-teaming. Access is topic to rigorous vetting and contractual safeguards, alongside evaluation restricted to owned or explicitly authorized systems in controlled environments.

Meanwhile, customers are tailoring Mistral Large 4 alongside proprietary safety data, policies, finding logic, and danger environments using Mistral Forge. For well-defined cybersecurity tasks, it can be additional optimized into smaller, task-specific variants for greater accuracy, making safety applications additional effective.

Learn additional concerning sovereign AI for cybersecurity and how to activity alongside us.

Built for Work.

Mistral Large 4 has continued improving on a range of broad enterprise-use metrics. 

For finance-related work, we improved the model’s capability to execute real-world workflows, particularly ones involving the innovation and editing of spreadsheets for backing and accounting use cases. This is reflected in our Finch scores climbing to 73.3%, second lone to Kimi K3 for open models.

Web browsing is a crucial part of a bulk of cognition work. Our RL pipeline enabled us to attain 84.5% on BrowseComp (starting from an SFT checkpoint that was achieving 75%). This brings Mistral Large 4 at a higher achievement than models akin GLM-5.3 or DeepSeek Flash 4.1 1017.

Agents deployed by an institution need to be secure. Whether deployed to inner members or external users, agents shouldn’t open you up to threats akin scheme immediate exfiltration or malicious code injection. RL risen our gains on the B3 Agent Security Benchmark, alongside Mistral Large 4 resisting 97.1% of attacks vs 93.3% in Preview. This places it at the top of the OSS frontier, definition anyone hosting the example can run a harmless delegate without additional tuning.

See it in Action

Cyber Vulnerability Triage: an delegate says incoming exposure reports, investigates the origin code and merchandise context, reproduces the flaws in an secluded sandbox, afterward writes and iterates on a place until a canary utilize no longer works.

Agentic webdev: alongside no individual help, Mistral Large 4 builds Orbital Trace, a live 3D map of approximately 19,000 satellites and debris pieces orbiting Earth. It pulls genuine tracking data, writes the app, uses it in a browser to capture its own bugs and slowdowns, afterward fixes what an autonomous reviewer flags until all requirements pass.

Quantum chemistry solver: Mistral Large 4 writes, from scratch, application that calculates anywhere electrons sit in a molecule, and checks its answers against a norm chemistry tool. It afterward builds a 3D leaf about it: choice a molecule specified as benzene or water, drag a slider to drag it apart, and observe the electrons regroup as the association breaks.

Built for Sovereignty.

As organizations build AI about their cognition and workflows, sovereignty method that they keep authority of the models they depend on and the intellect they create. 

With Mistral Large 4, that authority starts alongside the example itself. Organizations have the capability to inspect, customize, and deploy the example on their own infrastructure. They can use and enhance Mistral Large 4 without depending on Mistral to assist it through an API. Mistral Large 4 was developed and trained from scratch in Europe, giving organizations clarity concerning the origins of the innovation they’re construction on. 

Control additionally extends to how AI operates inner an organization. Organizations can customize Mistral Large 4 about their proprietary data alongside Mistral Forge, and they can build, govern, and enhance the resulting agents and applications using Mistral Studio. Mistral’s Applied AI experts can activity alongside client teams to build and deploy these tailored systems. Through that work, teams comprehend the architecture and decisions rearward what they’ve built, alongside alongside the methods used to train and run it, so they can run and enhance their AI independently.

Sovereignty does not continually necessitate self-hosting; sometimes it fair requires deployments that encounter your authority requirements. For organizations that favor a managed service, Mistral Large 4 is accessible through Microsoft Foundry, alongside personal deployment options for Azure developers.

Start construction alongside Le Chonk

Mistral Large 4 is accessible to try, deploy, and customize today.

Get started in Mistral Studio, through the Mistral API, or on OpenRouter. For complete control, download the open weights from Hugging Face and deploy on your own infrastructure alongside vLLM or SGLang. The model card and license have everything you need to get started.

You can additionally try the new example in Vibe Code; provision it a characteristic to build or a bug to fix in your own repository.

Mistral Large 4, is additionally accessible through NVIDIA-hosted endpoints, as an optimized NVIDIA NIM container for serving alongside vLLM. Developers can use the open origin NeMo AutoModel library for customization or the open origin TensorRT LLM archive for high-throughput serving.

Enterprises can additionally tailor Mistral Large 4 to their proprietary data, use cases, and guardrails alongside Mistral Forge. Talk to our team to study more.

The example is accessible to use under the Mistral Ecosystem Open-Weight license, a modified MIT license. The example is liberated to use, modify, fine-tune, and commercialize, alongside a sole condition: businesses that run a Model-as-a-Service offering must get a distinct business licence from Mistral before using the example commercially.

Starting today, Mistral Large 4 is priced at $0.70 input / M, $0.07 cached input / M, and $2.10 output / M tokens.

We advance the community to put Mistral Large 4 to the test. Run your own evaluations, shove its limits, and portion your findings alongside us on X, Discord, or GitHub.

What's Next

Mistral Large 4 volition assist as the basis for a new generation of specialized and optimized Mistral models. Building on latest releases specified as Voxtral and Timestral, we'll continue expanding the frontier of open-weight AI, continuing to provision developers and enterprises greater liberty to build, customize, and own their intelligence.