Newsbite analysis · AI, Data, Programming
The Decoder
Google's DiffusionGemma proves you don't need to train from scratch to build a text diffusion model
Google DeepMind introduced DiffusionGemma, a revolutionary model that shifts from traditional token-by-token text generation to processing blocks of 256 tokens concurrently. Utilizing a retrofitting method from the existing Gemma-4 model, this approach dramatically improves both speed and reasoning capabilities in tasks like Sudoku and structured outputs. Despite the advancements, the model indicates constraints in peak quality, particularly under heavy concurrent usage.
MonitorConfidence: MediumShort term horizon
Published 9 Aug 2026, 10:10 UTC4 min read
Lead image — The Decoder
Brief
NewsBite reading:Google's DiffusionGemma: Text Generation via Diffusion without Starting from Scratch
TL;DR
- DiffusionGemma processes blocks of 256 tokens in parallel, enhancing speed.
- Retrofitted from the Gemma-4 model, it requires less than 10% of the original training budget.
- Generates high-quality results in reasoning benchmarks, improving scores by an average of ten points.
- Maintains the option for word-by-word text generation, switching modes as needed.
- Model is utilized in applications such as multilingual speech recognition.
Interesting facts
- DiffusionGemma achieves about 1,500 tokens per second on Nvidia H100 accelerators.
- The model can solve nearly 85% of Sudoku puzzles correctly after minimal fine-tuning.
- Despite its speed, it can get stuck in repetition loops due to aggressive compute step reductions.