Short answer
Gemini Diffusion is an official Google DeepMind research model exploring diffusion for language and text generation. Google describes it as an experimental demo—not as a documented production API model. It is distinct from DiffusionGemma, a separate open experimental model built on Gemma 4 and Gemini Diffusion research. (deepmind.google)
The Google material cited here does not provide a production API model ID or endpoint for Gemini Diffusion. Do not infer access from another Gemini model’s API identifier or from a third-party listing.
What does Gemini Diffusion do?
Autoregressive language models generate text sequentially, predicting the next token from the preceding context. Google describes Gemini Diffusion as a research model that generates text by converting random noise into coherent text or code. Its page reports sampling speed and benchmark results, but those figures describe Google’s listed evaluations—not a guarantee of performance on your production workload. (deepmind.google)
For model selection, treat generation speed, task quality, and API access as separate questions. The diffusion approach alone does not establish that the model will be faster or better for your application.
What Google’s benchmark page reports
Google’s benchmark table compares Gemini Diffusion with Gemini 2.0 Flash-Lite. Among the listed results, it reports 89.6% vs. 90.2% on HumanEval, 76.0% vs. 75.8% on MBPP, and 15.0% vs. 21.0% on BIG-Bench Extra Hard, respectively. Google says the scores are pass@1 with no majority voting. Its methodology note specifies that the Gemini 2.0 Flash-Lite experiments used the AI Studio API, the gemini-2.0-flash-lite model ID, and default sampling settings. These are results for the stated evaluations, not a substitute for testing your own task. (deepmind.google)
The page also reports an average sampling speed of 1,479 tokens per second across its reported evaluations, excluding 0.84 seconds of overhead. Sampling speed alone is not end-to-end response time; consider overhead and measure the complete interaction your users will experience. (deepmind.google)
Gemini Diffusion vs. DiffusionGemma
| Name | What Google describes | Selection implication |
|---|---|---|
| Gemini Diffusion | An experimental research model exploring diffusion for language and text generation; Google describes an experimental demo. | Treat it as a research/demo option unless official API documentation provides a supported model ID and access method for your use. (deepmind.google) |
| DiffusionGemma | A separate open experimental model built on Gemma 4 and Gemini Diffusion research. Its documentation describes text diffusion and block-based generation. | Do not confuse this model’s documentation or availability with access to Gemini Diffusion through the Gemini API. (deepmind.google) |
The models are related but not interchangeable. Details about DiffusionGemma’s architecture or use do not automatically apply to Gemini Diffusion. (deepmind.google)
Can you use Gemini Diffusion through a production API?
The official material cited here does not document a production API model ID or endpoint for Gemini Diffusion. Google describes the model as available as an experimental demo. In its May 20, 2025 announcement, Google invited interested users to sign up for a demo waitlist; that historical announcement does not itself document present production API access. (deepmind.google)
Before building an integration, check Google’s current Gemini API model listing for an exact model name and the supported actions. Google documents models.list as a way to list models available through the Gemini API and inspect supported functionality. If the model is not listed for your account with the capabilities your application needs, do not treat the demo—or another model’s identifier—as API access. (deepmind.google)
How to choose for a production task
If you need a deployable API, choose only from models available to your account through the intended production channel. Then compare them on your own task:
- Confirm access: Find the exact model ID in the official API model listing and confirm it supports the action you need.
- Test task quality: Use representative prompts and expected outputs; assess correctness, consistency, and failure cases against your requirements.
- Measure full-request latency: Include overhead, not just sampling speed.
- Check applicable cost and limits: Consult current official API documentation for the relevant model and billing unit. Do not infer price or availability from a demo or third-party comparison.
Bottom line: Gemini Diffusion is an official experimental research model, and Google describes an experimental demo. The official material cited here does not document a production API model ID. DiffusionGemma is related but distinct. For production, verify a specific API listing and evaluate available models on your workload before selecting one.
Sources
Gemini Diffusion — Google DeepMind
Gemini Diffusion: Google DeepMind’s experimental research model