Google DeepMind quietly open sourced a “non-mainstream” model in early June:DiffusionGemma.Start with noise, gradually remove noise, and generate the entire text at once.
Diffusion model for text creation: how to do it?
Traditional LLM (such as GPT, Gemini) "writes" text word by word - each generated token depends on all previous tokens.
- Parallel generation: Place a large number of placeholder tokens on the "canvas" at one time, and then denoise and refine in multiple rounds of iterations
- bidirectional attention: The model can see all parts of the text simultaneously, allowing corrections and complex formatting to be inserted in real time
- Speed advantage: Reachable on a dedicated GPU 1000 tokens/second, 4 times that of the traditional autoregressive model
This idea was first unveiled as a demo at Google I/O 2025—at that time, the audience exclaimed when a model "instantly filled up the entire text" on the screen.
The advantages are outstanding and the disadvantages are also obvious
Speed is DiffusionGemma’s biggest selling point.
But Google also frankly announced its shortcomings: DiffusionGemma performs worse than standard Gemma 4 in all published benchmarks.
What does it mean for developers?
The real value of DiffusionGemma is not to “replace existing models”;Opening up a new text generation paradigm.
- Local real-time application: Requires low-latency client-side AI (game NPC dialogue, real-time subtitles)
- Batch text generation: Content production processes that require rapid generation of large numbers of drafts
- Research and exploration: AI researchers can explore new model design directions based on diffusion architecture
For the average user, DiffusionGemma is currently more of a technology preview than a practical tool.The future of AI text generation does not necessarily have to be just “one word and one sentence”..
The ecosystem of open source AI models is rapidly fragmenting. AI Dash ——Discover the best AI tools.