Skip to content
AI IntelligenceAug 9, 2026AI Intelligence
Article

Google DeepMind's DiffusionGemma shows that a text diffusion model can be built by retrofitting an existing model like Gemma...

Using less than 10% of the original training budget. It generates 256 tokens in parallel, reaching about 1,500 tokens per second.

Data Cube AI EditorialSource: The Decoder
01

Source Brief

Google DeepMind's DiffusionGemma shows that a text diffusion model can be built by retrofitting an existing model like Gemma 4, using less than 10% of the original training budget. It generates 256 tokens in parallel, reaching about 1,500 tokens per second.

02