Acyr Locatelli

A new method reduces KV cache size while maintaining high model performance.

BAM enhances MoE efficiency by integrating attention and FFN parameters.

Nexus combines efficiency, specialization, and adaptability in language model development.