DiffusionGemma
Google
A 26B MoE model (3.8B active params) that generates 256 tokens in parallel via discrete diffusion, delivering up to 4x speedup over autoregressive inference with bi-directional attention and iterative self-correction.
Parameters
26B MoE (3.8B active)
Context
256000
Modality
text->text
License
apache_2
Open source
Yes
Open DiffusionGemma in AIniverse
Compare versions, read real ratings, save to your stack.
Open in App