Mistral Releases Large-Scale Model Specialized for Coding

French AI company Mistral has released the weights of 'Large 4' (nicknamed 'Le Chonk'). This 1 trillion-parameter model was trained using only European domestic infrastructure, marking a new step in Mistral's open model strategy.

The 'number of parameters' in an AI model refers to the number of internal variables adjusted through model training, and generally, larger scale enables more complex processing. A trillion is the largest scale even at this point, and by being released as open source, it allows external researchers and developers to freely use and improve it—this is its major characteristic.

Mistral claims that Le Chonk outperforms open-source rival models in coding tasks. Independent evaluations have also shown that it has faster processing speed and relatively lower usage costs. Coding assistance is a high-demand field in AI applications, and this strength has practical value for businesses and developers.

On the other hand, a gap remains in comparison with top Chinese models. In benchmark (performance evaluation metric) scores, Le Chonk scored 26 points while DeepSeek's 'V4.1 Flash' and Z.ai's 'GLM-5.3' both recorded 32 points. A 6-point gap is not insignificant; however, given that Le Chonk was built within a self-contained European training environment, some view this as a notable achievement.

Over the past few years, the performance competition in AI models has been led by American and Chinese companies. Within this context, Mistral has continued developing models from a European perspective, holding significance as an option for users and organizations that prioritize data sovereignty and infrastructure transparency. From the perspective of EU AI regulation compliance, having a clearly defined training location can serve as a differentiation factor.

The release of Le Chonk is positioned as having a significant impact on the industry by broadening the choice of open models in the coding field. How much Mistral can narrow the performance gap with Chinese competitors in the future is likely to be the focal point in evaluating Mistral's progress. In the open model world, the release of weights itself is an asset, and improvements and applications by external communities are also anticipated; therefore, the value it holds cannot be measured by benchmark results alone.