Guide Labs debuts a new kind of interpretable LLM

The company open-sourced an 8 billion parameter LLM, Steerling-8B, trained with a new architecture designed to make its ...

Manifold-Constrained Hyper-Connections: The Architectural Breakthrough That Might Redefine LLM Training

If mHC scales the way early benchmarks suggest, it could reshape how we think about model capacity, compute budgets and the ...

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 ...