Cheaper & Faster & Smarter (TurboQuant and Attention Residuals)

About this article

Google TurboQuant This is a new compression algorithm. Every time a model answers a question, it stores a massive amount of intermediate data. The longer the conversation - the more expensive it gets. Result: compresses that data 6x+ with no quality loss, giving an 8x speed boost on H100s. No retraining required - it just plugs into an existing model Moonshot AI (Kimi) Attention Residuals The old way: each layer takes its own output and simply adds whatever came from the layer below. The new ...