Google retrofitted Gemma 4 into a diffusion model—and it now solves Sudoku

Google DeepMind didn’t build a new model from scratch. Instead, they took the existing Gemma-4-26B-A4B and, using less than 10% of the original training token budget, converted it into a diffusion model called DiffusionGemma. The technical report is out now, and it details a process that lets the model refine blocks of 256 tokens in parallel—hitting roughly 1,500 tokens per second on an Nvidia H100. That’s a fundamentally different approach from standard autoregressive models that have to commit to each token before seeing what comes next.

The secret sauce is a two-stage training process. First, the model learns to reconstruct noisy text blocks. Then comes a combined phase of reinforcement learning and sampler distillation, which Google dubs SD·RL. This merged approach is clever: reinforcement learning typically boosts answer quality, while sampler distillation slashes the number of compute steps needed. According to the report, this one-two punch raises scores on reasoning benchmarks by an average of ten points and nearly quadruples the number of tokens per compute step. A side effect is that DiffusionGemma’s answers run about 50% shorter, which further accelerates generation.

That parallel processing unlocks genuinely new behavior. In a math problem cited in the paper, the standard Gemma 4 model starts its answer with “-1,” works through the derivation, realizes the answer is “-25,” and then corrects itself. DiffusionGemma develops the reasoning and the answer simultaneously, fixing mistakes before the output is ever finalized. This bidirectional reasoning also shines in structured tasks. After minimal fine-tuning, the model solves nearly 85% of Sudoku puzzles correctly—a task the base model fails completely. For structured outputs like JSON or code repairs, it needs just two to three refinement steps because the input already constrains most tokens.

The tradeoffs are real, and Google is upfront about them. Absolute performance still trails the autoregressive base model. The model occasionally gets stuck in repetition loops, an artifact of aggressively reduced compute steps. The speed advantage also collapses under concurrent load; once about 32 simultaneous requests hit the model, standard language models catch up on throughput. Google explicitly labels DiffusionGemma as an experimental release, meant to accelerate research rather than dethrone autoregressive models. That’s a refreshingly honest framing in a field prone to overpromising. The model is already seeing real use—startup Interfaze is applying it to multilingual speech recognition, and another project is exploring interactive radiology report generation—suggesting that for specific, structured generation tasks, this retrofitting approach might be good enough right now.