NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·2h ago
Data Science
Simon Willison's Blog·Jun 10·5 min read
DiffusionGemma
TL;DR
There was a lack of accessible diffusion models for generating high-quality outputs. Google has released the open-weight Gemma model, named google/diffusiongemma-26B-A4B-it.
✦ Why It Matters
Engineers can leverage the Gemma model for efficient text generation in their applications.
Key Takeaways
- Access the Gemma model for free on NVIDIA's NIM cloud API.
- Experiment with token generation speeds exceeding 500 tokens per second.
- Utilize the open-weight model for your natural language processing tasks.
- Stay updated on advancements in diffusion models for future projects.
Full Summary
Last May, Google introduced an experimental Gemini Diffusion model, which generated excitement but was not followed by further announcements. Recently, this research has culminated in the release of the Gemma model, diffusiongemma-26B-A4B-it, under the Apache 2 license.
The model is hosted on NVIDIA's NIM cloud API, allowing users to access it for free. In testing, the model generated 2,409 tokens in just 4.4 seconds, achieving a speed of at least 500 tokens per second.
This performance indicates significant efficiency in token generation, which is crucial for applications in natural language processing. The open availability of this model could foster further innovation and experimentation in the AI community.