Mistral Large 4 enters public preview ahead of open-weight release

Model Architecture and Availability

Mistral AI announced the public preview of Mistral Large 4 on October 6, 2026. The model, referred to internally as ML4 or le Chonk, is described as a natively multimodal system with 1.05 trillion total parameters and 49 billion active parameters. It incorporates a one-million-token context window, positioning it as the company's largest and most capable offering to date.

The preview is accessible immediately through the Mistral Studio API, with the model weights scheduled for release by the end of October 2026. Training occurred on 3,800 NVIDIA Grace Blackwell GPUs within Mistral's European datacenters. The same infrastructure supports the public preview. A significant portion of the training data spanned more than 160 languages, encompassing every official language of the European Union.

The company has indicated that the model will be available across multiple regions, including a European deployment operated end-to-end by Mistral under European law. This setup aims to provide customers with greater control and autonomy, particularly relevant for sectors requiring sovereign AI solutions.

Performance in Cybersecurity and Agentic Tasks

Mistral Large 4 exhibits notable results on cybersecurity benchmarks. On the Artificial Analysis Cyber Index, it ranks among the top five models globally and leads open-weight models developed outside China. In one specific test involving reproduction and patching of a real vulnerability in open-source software, the model achieved an 82 percent score, the highest reported. It also solved 93 percent of challenges in Cybench, a benchmark comprising 40 exercises from security competitions.

These capabilities contrast with several leading closed models that score near zero on similar tests due to refusal mechanisms. The source notes that such refusals can impede legitimate vulnerability research and incident response. ML4's open-weight approach, combined with reduced moderation during red-teaming with cybersecurity leaders and vetted partners, is intended to mitigate provider-level constraints. Internal testing suggests additional utility in malware analysis, vulnerability prioritization, and detection rule writing.

Beyond cybersecurity, the model performs strongly on agentic coding benchmarks. It scored 61.7 percent on DeepSWE v1.1, 59.4 percent on SWE-Atlas-QnA, and 28.3 percent on Terminal-Bench 4.0. Its combined Coding Agent Index stands at 49.8 percent, ahead of certain comparable models. A blind human evaluation conducted with Surge AI rated ML4 Preview at 3.74 on a 1-5 scale for coding quality, placing it second among five models tested.

Multimodal Understanding and Scientific Applications

As a natively multimodal model, Mistral Large 4 demonstrates enhanced capabilities in processing images alongside text. It handles complex documents, charts, natural images, and specialized visual tasks relevant to engineering, manufacturing, and earth observation. The model integrates visual grounding with agentic functions, such as inspecting satellite imagery for disaster response or analyzing engineering drawings with precision.

Benchmark results indicate that on Dense 200, a visual grounding evaluation, ML4 achieved 42 percent compared to 41 percent for a leading closed model. The source highlights practical demonstrations involving dense natural scenes, technical drawings, PDF evidence retrieval, and geospatial image analysis. These features suggest potential value in time-sensitive or detail-oriented industrial applications.

In scientific and mathematical domains, the model is reported as state-of-the-art among open-weight systems on SciCode-Verified, which tests implementation of complex workflows in physics, mathematics, material science, and biology. It can generate a full Hartree-Fock simulation in a single pass, a multi-step chemistry task. Human evaluations indicate more precise and structured reasoning than a comparable model on internal STEM tasks in math and physics, supporting sustained domain-specific applied mathematics relevant to theoretical physics.

Knowledge Work, Finance, Legal Benchmarks, and Safety Considerations

Mistral Large 4 was evaluated on knowledge-work benchmarks relevant to professional environments. On AutomationBench, covering 657 business workflows across applications including Gmail, Google Sheets, Slack, and Salesforce, it scored 59.9 percent. It achieved 1,393 Elo on AA-Briefcase, which assesses long-horizon knowledge work producing deliverables such as spreadsheets, slides, and PDFs.

Third-party evaluation through vals.ai on representative legal and financial tasks found the model exceeding a leading closed model in both categories. On HarveyAI’s Legal Agent benchmark, it outperforms all open-source models. Performance on FinWorkBench, which involves creating and editing spreadsheets for real-life finance and accounting scenarios, further underscores its applicability to financial analysis requiring synthesis from multiple sources such as regulatory filings.

Safety evaluations place the model at the frontier for open-source systems. It resisted 93.3 percent of attacks on Lakera’s B3 AI Security Benchmark and demonstrated saturation on indirect prompt injection tests. On the KORA Benchmark, it recorded a score of 1.691, approaching the maximum of 2. The model shows a higher refusal rate on malicious cybersecurity prompts than other open-source models, despite strong benchmark performance in that domain. An internal human evaluation across coding, CAD, finance, mathematics, and physics preferred ML4 over a comparator in CAD and STEM tasks while performing comparably in finance and coding.

The company employs reinforcement learning at scale for post-training, adapting to the model's improving capabilities by training on its own attempt outcomes and progressively increasing task difficulty. This methodology underpins the model's current performance and is expected to support future specialized variants.

Strategic Context and Future Development

Mistral positions Mistral Large 4 as advancing AI sovereignty, particularly for European and enterprise users in regulated sectors such as cybersecurity, finance, and law. The open-weight release is framed as enabling self-deployment under organizational policies, addressing risks associated with dependency on external providers. Red-teaming with state authorities and partners continues until the weights are released.

The model serves as a foundation for forthcoming specialized and optimized versions. Additional details on architecture, benchmarks, and post-training methods are planned for disclosure prior to the weight release. Organizations in finance, engineering, manufacturing, logistics, pharmaceuticals, science, shipping, and the public sector contributed to its development through collaboration on training, customization, and reinforcement learning environments also offered to customers.

While the preview allows immediate testing and feedback, the full implications of open-weight deployment will become clearer after release. The reported benchmarks suggest competitive positioning against both open and closed models in targeted domains, though real-world enterprise integration will depend on factors including customization, latency, and compliance requirements. Continued refinement is acknowledged by the company as ongoing.

Source

Mistral AI published this record on 2026-10-06. Pennington Financial accessed it on 2026-10-09.