Google's New AI Writes Code All At Once

Описание: Google just open-sourced DiffusionGemma — a 26B model that writes code completely differently. Instead of predicting one token at a time, it generates the entire block at once by sharpening it out of noise. Over 1,000 tokens/sec on an H100. Apache 2.0, fully open weights.

Topics covered:
How diffusion language models work vs. standard autoregressive models
The speed numbers: 1,000+ tokens/sec on H100, 700+ on RTX 5090
Why parallel generation is naturally good at code (brackets, variables, consistency)
The Sudoku test: 0% → 80% after fine-tuning
The honest catch: overall quality is still below standard Gemma 4
What this unlocks: local speed, FIM edits, commercial use

BuildBeat covers every move in AI dev tools. Subscribe.