Researchers baked 3x inference speedups directly into LLM weights — without speculative decoding

Researchers from the University of Maryland, Lawrence Livermore, Columbia and TogetherAI have developed a training technique that triples LLM inference speed without auxiliary models or infrastructure ...

AI model delivers detailed 15-day Mediterranean Sea predictions in seconds

SeaCast is an innovative high-resolution forecasting system for the Mediterranean that harnesses AI to deliver faster and ...