BrestStormTeam

Mission:
We aim to efficiently train large-scale State Space Models (SSM) while significantly reducing infrastructure usage. Our goal is to minimize economic and environmental impacts without substantially compromising linguistic performance.

Model:
Tempest-LLM – an efficient language model based on Mamba2, leveraging advanced compression methods to achieve an encoding efficiency of 1.58 bits per parameter.

Training Approach:
Our model benefits from a balanced multilingual training strategy, ensuring equal proficiency in:

• 🇫🇷 French
• 🇬🇧 English
• 🇪🇸 Spanish

This multilingual training enhances linguistic versatility and cultural adaptability across different languages and contexts.

Impact:

• Economic: Reduced computational infrastructure leads to lower operational costs.
• Ecological: Lower power consumption and minimal infrastructure requirements decrease environmental footprint.
• Performance: Maintains robust linguistic accuracy and fluency despite compression and optimization.

Vision:
BrestStormTeam is committed to showing that linguistic AI technologies can be both powerful and sustainable, contributing responsibly to AI innovation.