DiffusionGemma: generates whole blocks of text at once, up to 4× faster on dedicated GPUs

DiffusionGemma：4倍速输出，整块文本同时生成

Google DeepMind released DiffusionGemma, an experimental open model that generates entire blocks of text at once instead of token by token. Output is up to 4× faster on dedicated GPUs, and the model can self-correct and format complex Markdown in real time. The post doesn't disclose parameter count, training data, or release timeline.

Why it matters: Google DeepMind drops an experimental open model that skips autoregressive decoding for 4x speed and self-correction — the approach is genuinely novel. But no params, training data, or release timeline disclosed, so we can't gauge how close this is to real use. Caps at 78.