Story perspectives

Google Releases Open-Source Gemma 4, 3× Faster Model

5/7/2026

Summary

Story summary

Google launched the Gemma 4 open-weight models with Multi-Token Prediction (MTP) drafters, promising up to a three-fold inference speedup.
Google re-licensed Gemma 4 under the Apache 2.0 license and made it runnable on TPU chips or quantized for consumer GPUs.
The MTP drafters use speculative decoding that halves token-generation wait time and can achieve up to a three-fold speedup.
Engineers Maarten Grootendorst and Olivier Lacombe say the speedup causes no loss in output quality.

1 of 1

Story summary

Google launched the Gemma 4 open-weight models with Multi-Token Prediction (MTP) drafters, promising up to a three-fold inference speedup.
Google re-licensed Gemma 4 under the Apache 2.0 license and made it runnable on TPU chips or quantized for consumer GPUs.
The MTP drafters use speculative decoding that halves token-generation wait time and can achieve up to a three-fold speedup.
Engineers Maarten Grootendorst and Olivier Lacombe say the speedup causes no loss in output quality.