Mistral AI's new model, Mistral Large 4, stands out with its features, 1 trillion parameter structure and success in legal processes. Details are in our news.
Mistral AI took its place in the technology world with the features of its new artificial intelligence model, Mistral Large 4, introduced under the code name Le Chonk. Although this model has a huge capacity of 1 trillion parameters, it activates only 49 billion parameters simultaneously using an efficiency-oriented Mixture-of-Experts architecture.
The model aims to provide the depth of a trillion-parameter structure at a lower computational cost by dynamically routing each token to private subnets. This structural approach allows the model to exhibit high performance when performing complex tasks.
Technical Architecture and Capabilities
Mistral Large 4 has a multi-mode structure that can process text and image data simultaneously. The model's attention layers have been optimized to natively support more than 160 languages, including all official languages of the European Union. Additionally, the model has special capabilities for determining and interpreting the coordinates of objects in complex images such as aerial and satellite photographs.
The internal structure of the model consists of two main elements called the attention layer and the Feed-Forward Network. While the attention layer establishes connections between words, the Feed-Forward Network stores the information of the model in the form of weights. In this way, the model reaches a capacity to manage complex and long-term agentic workflows. Especially in security-oriented studies, it can take an active role in tasks such as analyzing malware, recreating security vulnerabilities and applying patches.
During the training process, Mistral AI conducted a two-month study using 4,000 NVIDIA Grace Blackwell GPUs on European servers. The model utilized reinforcement learning pipelines to improve reasoning, coding, and human alignment capabilities. However, while the context window of the model was limited to 524,288 tokens, the output capacity was determined as 262,144 tokens.
Performance and Usage Areas
Looking at the performance of the model, it is seen that it achieved a score of 62 percent in the DeepSWE v1.1 benchmark test. In the evaluations made by Artificial Analysis, the model is defined as the smartest model developed outside the USA and China. Mistral Large 4 surpasses its competitors with its success in legislation and legal processes, especially in the Harvey's Legal Agent benchmark test.
On the pricing side, the cost of the model was determined as $1.13 per transaction for Intelligence Index tasks, according to Artificial Analysis data. In standard usage, a fee of $1.36 is charged for 1 million login tokens, $4.18 for 1 million login tokens, and $0.14 for 1 million cached login tokens. The technical depth and regulatory compliance offered by this model make it an important option, especially for corporate users. How do you think these technical features offered by Mistral Large 4 will affect the competition in the artificial intelligence market?
