Google turns Gemma into a faster text generator

Google DeepMind took their Gemma model and turned it into a diffusion model. This method uses only a small fraction of the original training budget. It is a much cheaper way to build new AI tools.

This new version creates 256 tokens at once instead of writing one word at a time. This change allows it to hit speeds of 1,500 tokens per second. It is a big step forward for generation speed.

The model is not perfect yet. It still performs worse than the original version on complex logic and reasoning tasks. It shows that speed and quality are still hard to balance.