Mistral Large 4 ‘Le Chonk’ challenges closed rivals with 1T-parameter open-weight MoE

Mistral Large 4 (Le Chonk) is a 1-trillion-parameter, natively multimodal mixture-of-experts model from Mistral AI that activates 49B to 52B parameters per token and is designed to compete with closed frontier models on cybersecurity, coding, agentic workflows, and visual grounding.

On Oct. 6, 2026, Mistral AI made Mistral Large 4 available as a public preview API on Mistral Studio, giving developers access to a model the company says “pushes the frontier of open-weight performance.” The release lands as European enterprises and public agencies look for sovereign alternatives to U.S. and Chinese foundation models, and as buyers increasingly judge models on measurable performance in security and agentic workloads rather than parameter counts. VentureBeat reported that the model, abbreviated ML4 and nicknamed Le Chonk, was trained from scratch over roughly two months on 4,000 Nvidia Grace Blackwell GPUs in Mistral’s own European data centers. The combination of European training infrastructure, planned open weights, and benchmark leadership is designed to make Le Chonk the default choice for European AI procurement.

Why does Le Chonk matter for Europe’s AI independence?

Mistral’s decision to train Le Chonk on its own infrastructure, rather than through a U.S. hyperscaler, gives European organizations an open-weight model whose training and hosting can remain inside the EU. The company says the model supports more than 160 languages, including all EU official languages, a requirement for government deployments and multilingual customer service across the bloc. According to Mistral’s launch announcement, the public preview began Oct. 6, 2026, via Mistral Studio, with open weights scheduled for the end of October. VentureBeat reported the training run took roughly two months on 4,000 Nvidia Grace Blackwell GPUs in Mistral’s own European data centers.

The timing is significant because enterprise AI procurement has shifted toward self-hosted and sovereign deployments, especially in regulated sectors such as defense, finance, and health care. An open-weight model that can be audited and run on-premises reduces dependence on API access to foreign models and lets organizations control where data flows. Mistral’s announcement explicitly frames ML4 as pushing “the frontier of open-weight performance,” a claim that independent evaluations will test between now and the open-weights release.

What benchmarks does Mistral Large 4 lead?

Mistral AI reports that ML4 scores 82% on CyberGym-E2E, a benchmark that requires a model to reproduce and patch a real vulnerability. That result places Le Chonk at the top of the AA Cyber Index among non-Chinese open-source software. The same evaluation campaign records 61.7% on DeepSWE v1.1 for software engineering and 59.9% on AutomationBench for agentic workflows. Those numbers are the headline evidence for Mistral’s claim that a trillion-parameter open-weight model can match closed rivals on tasks enterprises actually deploy, rather than on generic knowledge recall.

The cybersecurity result is especially relevant for enterprise buyers because most foundation models are not graded on end-to-end vulnerability remediation. CyberGym-E2E requires identifying a flaw, reproducing it, and producing a patch, a workflow that combines code understanding, tool use, and security domain knowledge. Mistral’s positioning suggests Le Chonk was optimized for the security-sensitive agentic work that defense, financial services, and critical infrastructure customers need.

How does Le Chonk fit the agentic enterprise stack?

Agentic AI is the most contested category in enterprise AI procurement because it determines whether a model can complete multi-step workflows without constant human intervention. Mistral’s 59.9% AutomationBench result is the company’s evidence that Le Chonk can drive those workflows, and the same agentic capability underpins the CyberGym-E2E result, where the model must navigate a security environment from detection to patch. For CIOs, the relevant question is not whether ML4 is a better chatbot but whether it can operate tools, parse logs, and finish jobs.

Visual grounding adds another layer. Because ML4 was trained natively multimodal, the vision encoder is part of the same model rather than an adapter bolted onto a text backbone. That matters for agentic workloads that involve screenshots, diagrams, and documents. A security analyst, for example, could ask the model to interpret a network diagram, read a vulnerability report, and generate a patch in a single session. Mistral’s docs list a 1.6B vision encoder and a 1M-token context, which gives the model room to hold both images and long codebases in context.

The combination of a 1M context window and multimodal input is what separates ML4 from earlier open-weight models in agentic deployments. Long-horizon tasks in software engineering and security often require a model to revisit dozens of files, read error logs, and update plans; a model with a short context loses that thread. Mistral’s benchmark numbers on DeepSWE v1.1 and AutomationBench suggest ML4 was optimized for that persistence, and the vision encoder extends it to tasks that require reading diagrams or screenshots.

What’s inside the 1-trillion-parameter architecture?

Mistral Large 4 uses a granular mixture-of-experts architecture. According to Mistral’s product documentation, the model has 1.05 trillion total parameters and 52 billion active parameters per token, alongside a 1.6 billion-parameter vision encoder. The launch announcement cites 49 billion active parameters, a discrepancy that likely reflects different counting conventions for vision-tower tokens or router overhead; both figures place ML4 in a different efficiency class from dense trillion-parameter models. The model accepts text and image input and generates text output across a 1 million-token context window.

The MoE design is central to the economics. A trillion total parameters sounds expensive, but because only 49B-52B are active per token, inference cost is closer to a 50B model than a 1T model. That is what makes open-weight deployment feasible for organizations that want sovereign AI without buying a massive GPU cluster. It also explains why Mistral can price the API at $0.68 per million input tokens and $2.09 per million output tokens during the launch window.

Specification	Value
Total parameters	1.05 trillion
Active parameters per token	49B-52B
Vision encoder	1.6B parameters
Input	Text and image
Output	Text
Context window	1 million tokens
Languages	160+, including all EU official languages
Launch API pricing	$0.68 / $2.09 per million input/output tokens
Open-weights release	End of October 2026 (around Oct. 27 per reports)

How does the pricing and rollout compare?

Mistral set launch pricing at $0.68 per million input tokens and $2.09 per million output tokens, a discount against list prices of $1.36 and $4.18 per million, according to Mistral’s docs. At the discounted rate, ML4 is priced below many closed frontier APIs, and the open-weights release will let organizations avoid API fees entirely by self-hosting. The public preview is available through Mistral Studio, so developers can test the model’s agentic and multimodal behavior before the weights are published.

The pricing strategy appears designed to build developer mindshare during the preview while preserving a paid API business after open weights arrive. Enterprises that validate ML4 on proprietary workloads during the discount period face switching costs if they later move to another model, and the option to deploy open weights on sovereign infrastructure reduces the commercial risk of an API dependency. VentureBeat’s report points to Oct. 27 for the open-weights release.

Enterprises comparing ML4 with closed frontier models will also look at operational details such as rate limits, latency, and fine-tuning options, which Mistral has not fully detailed. Open weights solve part of that problem by allowing self-hosting, but organizations that want a managed API need clarity on service-level agreements. The preview period is therefore as much a commercial test as a technical one.

What are the market and stakeholder implications?

For European AI buyers, Le Chonk is the first open-weight model outside China to claim leadership on cybersecurity benchmarks while also posting competitive coding and agentic scores. The AA Cyber Index placement gives procurement teams a metric they can cite when justifying a sovereign model over a closed U.S. or Chinese alternative. The fact that training happened in Mistral’s own European data centers strengthens data-governance arguments, because the training data and compute chain do not route through a third-country cloud.

The model also pressures closed frontier labs on price-performance. A 1T-parameter open-weight model with 49B-52B active parameters can be served on a modest cluster, and the 1M context window reduces the need for retrieval pipelines in some workloads. For enterprises, the practical comparison is not parameter count but cost per completed task; Mistral’s benchmark suite is designed to make that comparison favorable, even if independent labs have not yet reproduced the scores.

1. Watch for independent verification of CyberGym-E2E, DeepSWE v1.1, and AutomationBench scores between the Oct. 6 preview and the open-weights release.
2. Track the open-weights license terms, which will determine whether European government agencies can deploy ML4 in classified or regulated environments.
3. Monitor whether launch API pricing persists beyond the preview; list prices of $1.36 and $4.18 per million tokens are roughly double the discounted rates.
4. Assess enterprise adoption of ML4 for agentic security workflows, where the combination of 82% CyberGym-E2E and 59.9% AutomationBench is the model’s strongest pitch.

What do experts say about ML4?

Mistral co-founder and chief scientist Guillaume Lample told VentureBeat that “ML4 is at the frontier of open weight models.” The statement positions Le Chonk not as a European alternative that trails U.S. labs, but as a direct competitor to closed frontier systems. Mistral’s own announcement uses the same framing, describing the public preview as a release that “pushes the frontier of open-weight performance.”

ML4 is at the frontier of open weight modelsGuillaume Lample, co-founder and chief scientist, Mistral AI

The naming choice, Le Chonk, is an apparent nod to the model’s size, and Mistral embraced it in the launch post. That branding helps the model stand out in a release calendar crowded with incremental updates, but the substantive question is whether the benchmark scores survive independent replication. Lample’s quote, paired with Mistral’s documented scores, gives enterprise evaluators a falsifiable set of claims to test when the weights arrive.

What are the open questions before open weights?

The preview period leaves several questions open. Mistral’s benchmark scores are vendor-reported, and the most important validation will come when security researchers and enterprise teams run the open weights on their own workloads. The license terms will determine whether government agencies can deploy ML4 in regulated environments, and the exact release date around Oct. 27 will determine how quickly the open-source community can begin audits.

Technical questions also remain around the active-parameter discrepancy and the vision encoder’s performance outside benchmark distributions. Mistral’s docs list 52B active parameters while the launch post says 49B; neither figure changes the model’s class, but both will matter to engineers planning serving infrastructure. The 1M context window is another variable: long-context performance on real agentic tasks is often worse than benchmark scores suggest, and that will only be clear after hands-on testing.

What happens next?

The next milestone is the open-weights release at the end of October, around Oct. 27 according to VentureBeat. Once weights are public, security researchers and enterprise teams can audit the model, test it on internal benchmarks, and deploy it in sovereign environments. The outcome will shape whether Mistral can convert a strong benchmark release into durable European market share against both closed frontier labs and Chinese open-weight models.

Mistral has not announced a successor, a fine-tuning roadmap, or enterprise support tiers beyond the Studio API. The immediate test is reproducibility: whether independent teams can match the 82% CyberGym-E2E score, whether the vision encoder performs well outside benchmark distributions, and whether the 1M context window holds up on long-horizon agentic tasks. If those checks pass, Le Chonk will be the strongest sovereign alternative to closed frontier models available in Europe.