Mistral Large 4 Unveiled: 'Le Chonk', a Giant 1 Trillion Parameter MoE Model
Originally published in Turkish. Read the original →
Mistral AI crowned Europe's sovereign AI push with the 1-trillion-parameter Mistral Large 4. With an MoE architecture of 49 billion active parameters, a 1M context window and a plan to release open weights, a new era begins in the open-source world.
A historic announcement has been made that will shift the balance in the world of open-weight AI models. European AI pioneer Mistral AI officially introduced Mistral Large 4, its next-generation flagship model with the internal code name "Le Chonk". Standing out with its massive size of 1 trillion parameters (1T) in total, the model is positioned as one of the most comprehensive and capable multimodal AI systems ever trained in the open ecosystem.
Launched as concrete proof of Europe's technological independence and its sovereign AI vision, Mistral Large 4 directly challenges the top proprietary models in highly complex areas such as enterprise-grade coding, cybersecurity, law and financial analysis.
A 1 Trillion Parameter MoE Architecture and the Efficiency Balance
The Mixture-of-Experts (MoE) architecture that Mistral AI specializes in has been taken to its largest scale yet with Mistral Large 4. Although the model has a massive pool of 1 trillion parameters in total, it actively runs only 49 billion parameters for each token it generates (about 52 billion parameters including the input-output and embedding layers).
Thanks to this hybrid structure, Mistral Large 4 delivers the vast knowledge and reasoning depth of a giant 1-trillion-parameter system while keeping compute and inference costs at a level comparable to dense models in the 50-billion-parameter class.
A 1 Million Token Context Window and Native Multimodality
Beyond being a text-only model, Mistral Large 4 was designed from scratch with a natively multimodal architecture. The system, which can natively process images, charts, complex documentation and architecture diagrams, also supports a massive context window of 1 million tokens.
- Thanks to this context capacity, businesses can:
- Have technical specifications and legal contracts hundreds of pages long analyzed in one go,
- Feed large software codebases into the model in their entirety at the multi-repo level,
- Work through long conversation histories or multi-step agent workflows without any loss of context.
Trained in Europe on 3,800 NVIDIA Grace Blackwell GPUs
One of the most striking aspects of Mistral Large 4 is its training infrastructure and its focus on data security. The model was trained from scratch on a cluster of 3,800 NVIDIA Grace Blackwell GPUs that Mistral AI deployed in data centers located directly within Europe's borders.
This strategy offers a critical assurance especially for enterprise customers looking for a "sovereign AI" solution that fully complies with the European Union's AI Act (EU AI Act) and strict data privacy regulations. Financial institutions, healthcare organizations and public bodies can make sure their data is processed with the highest model capabilities without leaving their jurisdiction for that of third countries.
Open Weights Coming at the End of October
Staying true to its founding philosophy, Mistral AI is not locking Mistral Large 4 away as a closed API service only. The model is currently available for developers to test as a public API preview through Mistral Studio.
The company announced that it will release the model's open weights at the end of October 2026 so that the entire global research community and developers can download them to their own infrastructure. The distribution of the weights of a native 1-trillion-parameter MoE model in the open-source world marks a new turning point for the independent AI ecosystem running on local servers and private cloud clusters.
