Mistral Large 4 'Le Chonk' challenges closed US models with trillion-parameter open weights

Mistral Large 4 (ML4), nicknamed Le Chonk, is a 1-trillion-parameter natively multimodal mixture-of-experts model from Paris-based AI company Mistral AI, launched as a public preview Oct. 6, 2026, with open weights planned for release by the end of the month.

Le Chonk arrives at a moment when European governments and enterprises are pressing for AI systems they can audit, fine-tune and host on their own infrastructure. The model's 1 trillion total parameters are activated sparsely through a mixture-of-experts design, with Mistral citing 49 billion active parameters per token in its launch announcement and 52 billion in its documentation. That sparse activation is the engineering reason a model of this size can be served at costs closer to much smaller dense models. The preview is running on the same European datacenter infrastructure that trained the model, which Mistral says was 3,800 Nvidia Grace Blackwell GPUs.

The release is notable as much for what the model does as for what competing systems decline to do. Mistral reports that ML4 scored 82% on a benchmark that requires reproducing and patching real vulnerabilities in open-source software, the highest of any model the company tested. Claude Opus 5.5 and GPT-6 Astra, both closed frontier models, scored near zero on the same test, largely because of safety refusals. For enterprise security teams, that difference is the product: a model that can help defend networks without tripping over its own guardrails.

Why does Mistral Large 4 matter for Europe's AI sovereignty?

Mistral is among the most prominent open-weight AI developers headquartered in Europe, and ML4 is its largest release to date. The company trained the model from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own European datacenters, a deliberate departure from the common practice of renting North American compute. ZDNET's report on the launch says the training run used 4,000 Nvidia Grace Blackwell GPUs over two months. Either figure points to the same fact: the model was built and served without leaving European jurisdiction.

The sovereignty angle matters to government and regulated-industry buyers. France, Germany and other EU member states have signaled that they want AI models that can be audited, fine-tuned and hosted domestically, partly to reduce dependence on US providers subject to US law. Open weights, once released, let a government or enterprise run ML4 on its own hardware, its own network and its own data. Mistral says the model supports more than 160 languages, including all official EU languages, which extends its utility across the bloc's institutions and member-state agencies.

What is new in Mistral Large 4?

Mistral launched the public preview Oct. 6, 2026, via the Mistral Studio API. Open weights are planned for release by the end of October, with press reports specifying Oct. 27, after a safety review and testing with partners and authorities. The company describes ML4 as natively multimodal, meaning it was designed from the start to process images, charts and documents alongside text rather than bolting a vision encoder onto a text-only model. The documentation lists a 1.6 billion-parameter vision encoder and a 1 million-token context window, enough to hold a large codebase, a stack of legal contracts or a full technical manual in a single pass.

Mistral positions ML4 for coding, agentic workflows, cybersecurity, finance, law, manufacturing and visual grounding. For agentic use cases, the long context window and the software-engineering scores matter because agents that must navigate a repository, run tests and iterate need to retain substantial state. Mistral reports 61.7% on DeepSWE v1.1, a long-horizon software-engineering benchmark designed to measure sustained multi-file work rather than single-function generation. The company also says ML4 outperforms other open-weight models developed outside China on enterprise workloads, a qualifier that reflects the strength of Chinese open-weight labs in recent model releases.

Specification	Mistral Large 4 (Le Chonk)
Total parameters	1 trillion (1.05 trillion per Mistral documentation)
Active parameters	49 billion per launch announcement; 52 billion per documentation
Architecture	Natively multimodal mixture-of-experts with granular experts
Vision encoder	1.6 billion parameters
Context window	1 million tokens
Training hardware	3,800 Nvidia Grace Blackwell GPUs in European datacenters (ZDNET reported 4,000 GPUs over two months)
Languages	More than 160, including all official EU languages
Release status	Public preview Oct. 6, 2026; open weights planned Oct. 27, 2026
Distribution	Custom Mistral license after safety review

How does the model's technical architecture work?

ML4 is a mixture-of-experts model, a design in which a router selects a small subset of specialized subnetworks, or experts, for each token rather than activating the entire network. That is how a trillion-parameter model can run with only tens of billions of active parameters at a time. Mistral describes the architecture as granular, meaning the experts are smaller and more numerous, which the company says improves the trade-off between output quality and compute cost. The launch announcement cites 49 billion active parameters; the documentation lists 52 billion active parameters and 1.05 trillion total parameters. Both figures describe the same sparse-activation design.

Multimodality is native rather than added on. A 1.6 billion-parameter vision encoder lets ML4 accept images and documents, and Mistral says the model performs visual grounding — associating text with specific regions of an image — at a level that competes with closed frontier models. That capability matters in industrial inspection, document-heavy financial workflows and any setting where a model must both see a diagram and act on it. The 1 million-token context window also supports multimodal use cases that require passing entire image-heavy documents, such as annotated engineering drawings or scanned contracts, into a single prompt.

How well does Le Chonk score on benchmarks?

Highest of any model tested. Closed models Claude Opus 5.5 and GPT-6 Astra scored near zero, primarily due to refusals.Mistral AI

Mistral reports 82% on a cybersecurity benchmark involving reproducing and patching real vulnerabilities in open-source software, the highest of any model the company tested. On the Cybench cybersecurity benchmark, ML4 solved 93% of challenges, one of the highest scores for an open-weight model. On Lakera's B3 AI Security Benchmark, the model resisted 93.3% of attacks, a measure of how well it withstands attempts to jailbreak it into harmful behavior. The benchmark set is designed to show both offensive-adjacent capability in a defensive context and resistance to misuse.Mistral AI

On long-horizon software engineering, ML4 scored 61.7% on DeepSWE v1.1, a benchmark built around sustained multi-file coding tasks. The result trails the strongest closed models but sits among the best open-weight scores, and Mistral emphasizes that the model leads open-weight systems developed outside China on enterprise workloads. Because the weights are not yet public, independent labs have not been able to verify the numbers; the scheduled safety review ahead of the Oct. 27 release will be the first external checkpoint.

How does Le Chonk compare with closed US frontier models?

The comparison with closed models is where Le Chonk's cyber-defense positioning gets sharp. Claude Opus 5.5 and GPT-6 Astra scored near zero on the vulnerability benchmark, according to Mistral, not because they lack the capability but because their safety systems refuse the task. For a security team that wants to reproduce a published vulnerability to test its own defenses, a refusal is a dead end. Guillaume Lample, Mistral co-founder and chief scientist, framed the model as a defensive tool for enterprises and governments facing adversaries who abuse closed models:

The cyber defense capabilities will enable enterprises and governments to defend themselves against threat actors that are jailbreaking closed models to perform cyberattacks.Mistral AI

The asymmetry is deliberate. Closed frontier labs have been conservative about releasing models that can write working exploit code, even for defensive use, because the same capability can be misused. Mistral's bet is that an open-weight model with strong defensive posture — including the 93.3% resistance to jailbreak attacks on Lakera's B3 benchmark — is a net positive for security. Lample's competitive claim, reported by VentureBeat, was direct: 'ML4 is at the frontier of open weight models.'

What are the market and stakeholder implications?

Mistral is aiming ML4 at buyers who need model sovereignty: banks with regulatory obligations around data residency, manufacturers running industrial vision systems, law firms handling privileged documents, and government agencies that cannot send sensitive material to third-party APIs. The open-weights release, planned for Oct. 27 after review with partners and authorities, would convert those buyers from API customers into operators of their own infrastructure. In the interim, the Mistral Studio preview and Mistral Forge offer evaluation and fine-tuning paths.

1. Run the open weights on internal infrastructure once they are released, keeping data in-region for regulated industries.
2. Evaluate the preview through the Mistral Studio API for coding, agentic and multimodal workloads.
3. Use Mistral Forge to fine-tune and deploy domain-specific variants for finance, law or manufacturing.
4. Deploy ML4 for defensive cyber work such as vulnerability reproduction, patching and attack-surface review.
5. Build agentic systems that use the 1 million-token context window for long-horizon software engineering and document analysis.

The commercial stakes are considerable for a company competing against much larger US labs. Open-weight models tend to commoditize the API layer, pushing revenue toward infrastructure, fine-tuning and enterprise services, and ML4's benchmark story gives Mistral a differentiated pitch: a model that is both open and defensively strong. That pitch lands at a moment when enterprises are increasingly wary of the cost and governance of closed frontier APIs, and when European policymakers are actively funding and procuring domestic AI capability.

What do experts say about the model?

Lample's public comments frame ML4 as both a technical and a geopolitical product. In ZDNET's interview, he emphasized cyber defense as the model's headline capability for governments. In VentureBeat's coverage of the launch, he staked out the competitive position: 'ML4 is at the frontier of open weight models.' Mistral has not named independent benchmark auditors, so the public record currently rests on its own evaluations and on press reporting such as ZDNET's account of the training run. The Oct. 27 weight release, when third parties can inspect and reproduce the model, will be the moment the claims face independent scrutiny.

ML4 is at the frontier of open weight models.Guillaume Lample, Mistral co-founder and chief scientist

The practical significance of that claim depends on the license. Mistral has said the weights will be distributed under a custom license, and the terms will determine how freely enterprises can fine-tune, host and commercialize the model, and whether derivative models can be published. For governments, the relevant question is whether the license permits deployment on sovereign infrastructure with no ongoing connection to Mistral's services. The company has not yet published the full license text.

What happens next?

The near-term timeline is fixed. Weights are planned for release by Oct. 27, 2026, after a safety review and testing with partners and authorities, under a custom Mistral license. Whether that release happens on schedule, and with what conditions, will be the first test of Mistral's claims. The second test is external: once weights are public, security researchers and benchmark groups can verify the 82% vulnerability-patching score, the 93% Cybench result and the 93.3% jailbreak resistance on infrastructure outside Mistral's control.

For the frontier-model landscape, the release sharpens a division already visible across the industry: US closed labs increasingly gate their strongest systems behind APIs and safety policies, while open-weight labs in Europe and China ship auditable models that close the capability gap. Le Chonk's cyber-defense scores are the clearest example yet of a model whose value to buyers comes directly from what closed competitors refuse to do. Whether that posture survives external safety review — and whether the weights ship on schedule — will determine whether Mistral's sovereignty pitch translates into enterprise and government adoption.