Google AI Updates: August 10, 2026

1. Google’s DiffusionGemma Retrofits Gemma Into a Text Diffusion Model

Google. Google DeepMind released DiffusionGemma, a text diffusion model that refines blocks of 256 tokens in parallel, reaching roughly 1,500 tokens per second, built by adapting the existing Gemma-4-26B rather than training from scratch. The team used less than 10 percent of the original training token budget across two stages, supervised fine-tuning followed by combined reinforcement learning and sampler distillation. The result shows practitioners can convert strong autoregressive models into diffusion generators cheaply, unlocking bidirectional reasoning and self-correction that suit coding and structured output tasks. Source