Google Releases Multi-Token Prediction Drafters for Gemma 4 — Up to 3x Faster Open-Source Inference

Google / Hacker News·high signal

Google released open-source Multi-Token Prediction (MTP) drafters for the Gemma 4 model family, enabling up to 3x faster inference by pairing a heavy target model (e.g., Gemma 4 31B) with a lightweight drafter that predicts several future tokens in parallel. The drafters are available under Apache 2.0 on Hugging Face and Kaggle, compatible with transformers, MLX, vLLM, SGLang, and Ollama — meaning developers can deploy faster inference on personal machines immediately. Currently at 645 points and 315 comments on HN, this is the highest-engagement AI story on the front page.