Search
On June 10 (local time), Google introduced an experimental open-source model named DiffusionGemma. Leveraging a text diffusion framework, this model is capable of generating text at speeds up to four times faster than conventional autoregressive large language models when run on specialized GPUs. It is made available under the Apache 2.0 license. Google clarified that DiffusionGemma is primarily designed for researchers and developers, noting that its overall output quality falls short of that of the standard Gemma 4 model. For production settings, Google still advises using the standard Gemma 4. Furthermore, the speed superiority of DiffusionGemma is most noticeable in local and low-concurrency inference situations, while its advantages are less significant in high-concurrency cloud deployment scenarios.
-
C114 Communication Network -
Communication Home
