Google DeepMind researchers developed DiffusionGemma, a text diffusion model that leverages an existing model, Gemma 4, to significantly reduce training costs. By parallelizing token generation, DiffusionGemma achieves a higher output rate than its predecessor, reaching 1,500 tokens per second. However, the model's performance still lags behind the original autoregressive model, particularly in tasks requiring reasoning. The development demonstrates a more efficient approach to building text diffusion models, which could have implications for large-scale natural language processing applications.