Post #6862

⚡️Google just found a way to make Gemma-4 generate text nearly 3× faster Instead of predicting words one token at a time like most transformer models, Google introduced Multi-Token Prediction (MTP) drafters for Gemma-4. Here’s the trick: A smaller auxiliary model “drafts” several tokens ahead, while the main model checks the entire batch in a single pass. If one token is wrong, it gets corrected and generation resumes from that exact point. That means fewer expensive passes through the main model and a huge speed boost without hurting output quality. Google says the new system delivers around 3× faster generation, and they’ve open-sourced all the drafters so developers can already start experimenting with it. AI labs are no longer just competing on intelligence. The next battleground is speed. @aipost🏴