Mistral Large 4, a strong coding model but still behind Chinese models
In brief
- Mistral releases Large 4, a 1-trillion-parameter model trained in Europe and nicknamed "Le Chonk".
- It outperforms other open models in coding but still trails Chinese models in overall performance.
- The model is available under a modified MIT license and accessible via Hugging Face and Microsoft Foundry.
Mistral has released the weights for its largest model yet, Mistral Large 4, codenamed "Le Chonk", trained entirely in Europe. The company claims it outperforms other open models in coding while remaining fast and cost-effective. However, it still lags behind top Chinese models, scoring 26 points compared to 32 for DeepSeek V4.1 Flash and GLM-5.3.
The trillion-parameter model handles both text and images as input and generates text and code. It requires significant hardware, such as eight Nvidia B300 GPUs for the standard version. After three weeks of public preview, Mistral refined the model using reinforcement learning, boosting its coding and cyber defense capabilities. Independent tests show notable progress, though gaps remain with closed models like Claude Opus 5.5.
Mistral positions Large 4 as a "sovereign" AI model, trained in Europe and available under a modified MIT license. It is accessible via Hugging Face and platforms like Microsoft Foundry. Despite strong coding and cybersecurity performance, it struggles to compete with leading Chinese or US models, while remaining competitive in cost and speed.
What it changes for you
- CybersecurityCybersecurity teams can test Large 4 for defense tasks, but must monitor its limitations compared to Chinese and US models.
- EducationEducational institutions can use Large 4 for programming projects, but its cost and hardware requirements limit accessibility.
This summary gives you the essentials. The full article, its figures and quotes are on The Decoder.
Read the original articleAI-generated summary of the source, checked by rules. How we write · Report an error