Mistral Large 4 unveiled: Europe's trillion-parameter model enters preview

French AI company Mistral announced the public preview of its largest model, Mistral Large 4 (ML4), on 6 October 2026. Codenamed "Le Chonk," the model has been made available as an API through the company's Mistral Studio platform. Model weights are planned to be released at the end of October; until then, ML4 is listed as a proprietary model since its weights haven't been shared.

According to Mistral, ML4 is natively multimodal and has a total of 1 trillion parameters, 49 billion of which are active. The company claims the model outperforms even proprietary rival models in cyber defense, manufacturing, and finance.

What do independent benchmarks show?

On Artificial Analysis's Intelligence Index, which combines ten different domains, ML4 scored 38 points. This marks a major leap compared to the previous model Mistral Large 3's 9 points and the recently released Mistral Medium 3.5's 14 points. With this score, ML4 surpasses Z.ai's GLM-5.2 model, but the leader, Claude Opus 5.5 (Max), remains far ahead with 58 points. In coding, ML4 scored 49.8 percent on the Artificial Analysis Coding Agent Index, putting it ahead of Deepseek V4 Pro and Qwen3.8 Max. In Surge AI's blind evaluation, ML4 ranked second among five models with 3.74/5 points; first place went to Claude Opus 5 with 4.22 points.

Where does the cybersecurity claim come from?

Mistral devoted much of its announcement to cybersecurity capabilities. According to the company, ML4 achieved 82 percent success in a test of reproducing and patching a real open-source software vulnerability, the highest score among all models tested. However, Mistral notes that models like Claude Opus 5.5 and GPT-6 Astra completely refused this task due to security filters, meaning the test measures provider policies as much as model capability. Mistral also reports high refusal rates against malicious cyber prompts, stating it blocked 93.3 percent of attacks on Lakera's B3 AI Security Benchmark.

What does it offer for technical drawing and image analysis?

Mistral claims ML4 can analyze documents, charts, gigapixel satellite imagery, and technical drawings, and can zoom in and examine details on its own. On the Dense 200 benchmark, ML4 scored 42 percent, slightly ahead of GPT-6 Astra's 41 percent. The company also states that human evaluators preferred ML4 on STEM and CAD tasks—a notable detail for teams working with architectural and engineering documents.

What does this mean for architecture and archviz teams?

The model's performance on technical drawing and CAD tasks suggests potential future use in document analysis, quantity takeoff checks, or automated scanning of technical sheets in architectural projects. However, the model is still in preview and considered proprietary; it would be premature for offices to plan on-premise deployment or integration before the weights are released and licensing terms are finalized. The emphasis on European hosting and data sovereignty may appeal to architecture offices in Europe from a data privacy standpoint, but given the performance gap, closed models like Claude and GPT-6 Astra are likely to remain preferred for daily work.

Mistral stated it trained ML4 from scratch in its own European data centers using 3,800 Nvidia Grace Blackwell GPUs. The model's context window extends up to 1 million tokens; according to documentation, the architecture features a fine-grained mixture-of-experts structure and a 1.6 billion-parameter vision encoder. During the preview period, pricing has been set at $0.68 per million input tokens and $2.09 for output tokens; cached inputs are billed at $0.07. The documentation also lists standard prices double these rates. The expansion is funded by the company's $3 billion Series D round.