DiffusionGemma: Google Converts Gemma 4 Into a Diffusion Model for Under 10% of Training Cost

Google DeepMind took the existing 26-billion-parameter Gemma 4 and converted it into a diffusion model for less than 10% of the original training budget. DiffusionGemma generates 256 tokens simultaneously — not one at a time like a standard LLM — and delivers around 1,500 tokens per second on an Nvidia H100. Quality is slightly below the original, but reasoning tasks improved by 10 points, and Sudoku puzzles are solved 85% of the time — versus zero for the base model.

Until now, diffusion-based text models required training from scratch — a barrier that kept all but the largest players away. Now that any existing model can be repurposed for a fraction of the budget, specialized variants will become cheaper and more diverse. This reshapes the economics of the race for generation efficiency.