Google DeepMind converted its existing Gemma-4-26B-A4B model into a text diffusion model called DiffusionGemma, using less than ten percent of the original training token budget instead of training from scratch. The two-stage process first teaches the model to reconstruct noisy text blocks, then combines reinforcement learning and sampler distillation in a phase called SD·RL. This approach raises quality on reasoning benchmarks by an average of ten points while nearly quadrupling the number of tokens per compute step, and DiffusionGemma's answers run about 50 percent shorter. Unlike autoregressive models, DiffusionGemma develops answers and reasoning in parallel, allowing it to correct mistakes before finalizing output. After minimal fine-tuning, it solves close to 85 percent of Sudoku puzzles correctly, while the base model fails entirely. Structured outputs like JSON or code repairs finish after just two to three refinement steps. DiffusionGemma retains its original word-by-word text generation ability, letting users switch between modes. It trails the autoregressive Gemma 4 on quality benchmarks but leads on output speed at about 1,500 tokens per second. Absolute performance falls short of the base model because DiffusionGemma was retrofitted after the fact, its subsequent training was relatively short, and SD·RL prioritized speed over peak quality, with architecture and training data carried over from the original.