Mistral Unveils Large 4, a European Model with One Trillion Parameters

Mistral unveils a preview version of Mistral Large 4, presented as multimodal and trained in Europe. The model shows significant progress in aggregated rankings but remains behind closed leaders, while the company focuses on cybersecurity and sovereign deployment. The weights and technical details are expected to be announced by the end of October.

Publication of Weights, Ongoing Training, and Proprietary Scope

Mistral Large 4 is currently considered proprietary, with the weights not yet published. In the meantime, the company is testing the model with security companies, verified partners, and government agencies, providing them with a version that has reduced moderation and extended cyber capabilities. Mistral indicates that the reinforcement learning cycle supporting the preview is still ongoing and shows no signs of plateauing, with improvements expected in the coming weeks. The publication of weights and details on architecture, licensing, and post-training methods is anticipated by the end of October.

European Capabilities: GPU, Data Centers, Languages, and Roadmap

According to Mistral, the training of Mistral Large 4 was conducted entirely on 3,800 Nvidia Grace Blackwell GPUs installed in its own data centers located in Europe, and the preview also operates on this same infrastructure. The company announces an independently operated European version, without reliance on other providers, compliant with European Union regulations, and specifies that the data used for training encompasses over 160 languages, including all official EU languages. The training involved players from finance, industry, logistics, pharmaceuticals, maritime, and the public sector, utilizing the environment offered through Mistral Forge, with approximately 3,000 GPUs for reinforcement post-training and a throughput of about 33 billion tokens per day. The increase in computing power is supported by a €3 billion round presented as a European record, and an $830 million loan taken in March for a center near Paris, aiming for 200 MW of capacity in Europe by the end of 2027. Meanwhile, the company is redirecting its offering towards professional uses, having renamed its chatbot Le Chat to Vibe in May and rebuilt it as a work tool.

Cybersecurity: Vulnerability Testing, Targeted Refusals, and On-Site Execution

According to the company, Mistral Large 4 ranks among the top five models in the world in a cyber index, leading among open-weight models developed outside of China. In a test requiring the reproduction of a real vulnerability in open-source software and its correction, the model achieved 82%, the highest score, while Claude Opus 5.5 and GPT-6 Astra scored close to zero as they refused the task. This exercise reflects both the policies of the providers and the technical capabilities. Mistral argues that proving the existence of a flaw falls under software defense and that security filters of closed models block this work, adding that attackers circumvent these systems and that loss of access to a provider during an incident can become a risk. The model is announced as deployable in a private cloud or on-site. On malicious prompt sets such as JailbreakBench, StrongREJECT, and AgentHarm, the company highlights refusal rates superior to other open models, while not explaining how the distinction between legitimate research and attack preparation is made. In the Lakera B3 benchmark, the model would block 93.3% of attacks. In May, Arthur Mensch alerted a French parliamentary committee about the risk of dependence on American models for cybersecurity, particularly noting that French military codebases should not be scanned by Mythos from Anthropic, asserting that Mistral's models could encounter the same vulnerabilities.

Aggregated Benchmarks: Notable Progress but Persistent Gap at the Top

In the Intelligence Analysis Index, which compiles ten benchmarks, Mistral Large 4 scores 38 points, compared to 9 for Mistral Large 3 and 14 for Mistral Medium 3.5, surpassing Z.ai's GLM-5.2. At the top, Claude Opus 5.5 (Max) scores 58 points, while Mistral's model remains behind the closed offerings from Anthropic, OpenAI, and Google, as well as several Chinese open models. The model is close to the top in scientific coding, although GPT-6 Astra and Qwen3.8 maintain higher scores. In the Coding Agents Index, it reaches 49.8%, ahead of Deepseek V4 Pro and Qwen3.8 Max. A blind evaluation conducted with Surge AI places it second out of five with a score of 3.74/5, ahead of GLM-5.3 and Kimi K3, but behind Claude Opus 5 (4.22). Human evaluations reported by Mistral favor the model in STEM and CAD tasks, with performance deemed comparable in finance and coding. On automated streams, it ranks slightly behind GLM-5.3 and ahead of Kimi K3 and Deepseek V4 Pro.

Vision and Documents: Close Results and Third-Party Evaluations

Mistral presents Mistral Large 4 as a native multimodal model and claims an advantage in visual anchoring over closed models. The company announces a marked improvement in image analysis, with support for documents, graphics, gigapixel satellite images, and technical plans, including zoom and autonomous verification operations. On Dense 200, the gap remains narrow, with 42% for Mistral Large 4 compared to 41% for GPT-6 Astra. According to Vals.ai, the model also outperforms GPT-6 Astra on legal and financial tasks.

Product, Architecture, and Pricing During the Preview

The API is accessible via Mistral Studio. Mistral specifies that 49 billion parameters are used simultaneously and presents a mixture-of-experts architecture with fine granularity comprising a total of 1.05 trillion parameters, a vision encoder of 1.6 billion, and a context window of one million tokens. During the preview phase, the communicated rates are set at $0.68 for each million input tokens and $2.09 for a million output tokens, while cached inputs are charged at $0.07, with documentation also noting doubled rates ($1.36, $4.18, and $0.14). Mistral plans to leverage this model to develop a new generation of specialized models, while highlighting benchmark performances in various fields and, according to aggregated benchmarks, a position as the best open-weight model in the United States or Europe.