Gemini Diffusion
a year ago
- #AI
- #Diffusion Models
- Gemini Diffusion is Google's first LLM using diffusion instead of transformers.
- Diffusion models refine noise step-by-step, enabling faster and more coherent outputs.
- Gemini Diffusion excels in tasks like editing, math, and code due to its iterative process.
- The model boasts speeds of 857 tokens/second, generating interactive HTML+JavaScript quickly.
- Performance is comparable to Cerebras Coder running Llama3.1-70b at 2,000 tokens/second.
- Google claims it matches Gemini 2.0 Flash-Lite's performance at 5x the speed.
- Prior to this, Inception Mercury was the only commercial-grade diffusion model available.