Google launched the open-source model DiffusionGemma.

Google released an experimental open-source model called DiffusionGemma on June 10th local time. It uses a text diffusion architecture and achieves a maximum 4x speedup in text generation on dedicated GPUs compared to traditional auto-regressive large language models. The model is released under the Apache 2.0 license. Google states that DiffusionGemma is positioned as an experimental model for researchers and developers, and overall output quality is lower than the standard Gemma 4, so it is still recommended to use the latter in production environments. The speed advantage of this model is mainly seen in local and low-concurrency inference scenarios, with limited advantage in high-concurrency cloud deployments.