What happened
Mistral AI announced the public preview of Mistral Large 4 (ML4), a 1 trillion-parameter natively multimodal model with 49 billion active parameters. The model is currently accessible via the Mistral Studio API, with full open- release planned for the end of the month. Mistral states that ML4 was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European datacenters. The company reports that ML4 achieves state-of-the-art performance among open models in enterprise workloads, including cybersecurity, finance, and law, and surpasses certain closed models in visual grounding tasks. The launch coincides with Mistral's €3 billion Series D funding round, which is being used to scale compute capacity in European datacenters.
Mistral AI has launched a public preview of Mistral Large 4, internally referred to as 'le Chonk' or ML4. The model is a 1 trillion-parameter natively multimodal architecture with 49 billion active parameters. It is currently available via the Mistral Studio API, with the full open- release scheduled for the end of the month. Mistral states that the model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs located in its own datacenters in Europe.
The company claims ML4 demonstrates exceptional performance across coding, agentic workflows, and multimodal understanding. Mistral reports that ML4 is competitive with the strongest open-source models globally and significantly outperforms open- models developed in the US or Europe. In specific enterprise verticals such as cybersecurity, finance, and law, Mistral asserts the model is state-of-the-art among open models. In visual grounding tasks, such as the Dense 200 benchmark, Mistral claims ML4 surpasses GPT-6-Astra with a score of 42% compared to 41%.
A key differentiator highlighted by Mistral is the model's cybersecurity capability. On the Artificial Analysis Cyber Index, ML4 ranks among the top five models globally and leads open- models developed outside China. Mistral notes that ML4 scores 82% on a test requiring the reproduction and patching of a real vulnerability, a task where several leading closed models, including Claude Opus 5.5 and GPT-6 Astra, score near zero due to refusal mechanisms. Mistral argues this allows defenders to perform legitimate vulnerability research and incident response without being blocked by provider-level safety filters.
The model is also positioned for scientific and professional workloads. Mistral reports that ML4 is state-of-the-art on SciCode-Verified among open- models and can generate complex simulations, such as a Hartree–Fock simulation, in a single shot. In coding benchmarks, ML4 scores 61.7% on DeepSWE v1.1 and 59.4% on SWE-Atlas-QnA. A blind human evaluation by Surge AI ranked ML4 second among five models for coding quality, behind only Claude Opus 5.
This launch is part of Mistral's broader strategy following its €3 billion Series D funding round, the largest equity round ever raised by a European technology company. The capital is being used to scale compute capacity in European datacenters. Mistral states that the run behind the preview is still in flight and that the model is expected to improve further as training continues on expanded infrastructure.
Why it matters
The release of a 1 trillion-parameter open- model from a European company represents a significant shift in the global AI landscape, offering an alternative to US-dominated closed models. By providing open weights, Mistral enables organizations to deploy the model on private clouds or on-premise, addressing sovereignty and data privacy concerns critical for sectors like finance, law, and cybersecurity. The model's specific emphasis on cybersecurity capabilities, including the ability to reproduce and patch vulnerabilities without the refusal mechanisms common in closed models, offers a practical tool for defenders who need to audit and secure software without being blocked by safety filters. This development reinforces the trend toward open-weight models as viable, high-performance options for enterprise and critical infrastructure applications.
The release of Mistral Large 4 as an open- model provides a significant alternative to closed-source frontier models, particularly for organizations with strict data sovereignty or privacy requirements. By allowing self-deployment on private clouds or on-premise, ML4 enables customers to maintain control over their AI infrastructure, which is critical for sectors like finance, law, and government.
The model's specific focus on cybersecurity capabilities addresses a practical gap in current AI tools. Closed models often refuse to perform tasks that involve reproducing vulnerabilities or analyzing malware, citing safety concerns. Mistral argues that these refusals can hinder legitimate defensive work, such as incident response and vulnerability research. By providing a model that can perform these tasks without the same level of refusal, Mistral offers a tool that security teams can use to audit and secure their systems more effectively.
The training of a 1 trillion-parameter model entirely within European datacenters on European infrastructure represents a milestone for the region's AI sovereignty. It demonstrates that Europe can develop and deploy frontier-scale AI models independently of US or Chinese infrastructure, which has geopolitical and economic implications for the global AI landscape.
The model's performance in multimodal and agentic tasks, such as analyzing satellite imagery for disaster response or managing complex business workflows, suggests it can be applied to a wide range of practical, high-stakes scenarios. This broadens the utility of open- models beyond simple text generation to complex, real-world problem-solving.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
crm_get_transaction(id='4092').Which component of an AI application is the machine-learning model itself?
What to watch next
Monitor the release of the full open weights at the end of the month to verify the model's actual performance against the reported benchmarks. Watch for independent third-party evaluations that may differ from Mistral's internal or partner-reported scores. Observe how the model's cybersecurity capabilities are received by security researchers and whether the lack of strict refusal mechanisms leads to any misuse or regulatory scrutiny. Track Mistral's expansion of its European datacenter infrastructure as part of its Series D investment.
The full release of the open weights at the end of the month will be the next critical step. Independent researchers and developers will be able to test the model's actual performance, verify the reported benchmarks, and explore its capabilities in ways that may differ from Mistral's internal evaluations.
The model's cybersecurity features, particularly its ability to perform offensive-style tasks like vulnerability reproduction without refusal, may attract scrutiny from regulators and security experts. There is a potential risk that these capabilities could be misused by malicious actors, and the balance between defensive utility and safety will be a key point of discussion.
Mistral's expansion of its European datacenter infrastructure, funded by the €3 billion Series D, will determine its ability to sustain and scale the training and deployment of such large models. The pace of this expansion will impact the company's long-term competitiveness and its ability to deliver on its promises of sovereign AI.
The model's performance in scientific and specialized domains, such as physics and chemistry, will be closely watched by the research community. If ML4 can reliably assist in complex scientific workflows, it could accelerate research in these fields and establish open- models as essential tools for scientific discovery.