At Google I/O yesterday, the company unveiled Gemini Diffusion, its latest innovation in language models. Unlike traditional autoregressive models, which generate text sequentially, one word at a time, Gemini Diffusion uses a diffusion technique to produce content. This method, inspired by image generation models like Imagen and Stable Diffusion, allows the model to refine noisy data iteratively, leading to faster and more coherent output. Google explains that traditional autoregressive models can be slow and may struggle with maintaining high quality and consistency in their generated text. On the other hand, diffusion models work by gradually refining noise into meaningful content. This iterative process enables quick error correction and enhances the model's ability to perform tasks such as editing, especially in contexts involving mathematics and coding. I had the opportunity to test Gemini Diffusion after making it through the waitlist, and the speed is indeed impressive. When prompted with "Build a simulated chat app," the model generated an interactive HTML+JavaScript page in single-digit seconds, delivering 857 tokens per second. This performance is reminiscent of the Cerebras Coder tool, which runs Llama3.1-70b at around 2,000 tokens per second. While there are currently no independent benchmarks available, Google’s landing page claims that Gemini Diffusion offers the performance of their Gemini 2.0 Flash-Lite model, but at five times the speed. Given this assertion, it seems Google is confident that Gemini Diffusion maintains a high level of quality comparable to their less expensive models. Before Gemini Diffusion, the only commercial-grade diffusion model I had encountered was Inception Mercury, released in February. However, it’s important to clarify that diffusion doesn’t replace transformers entirely; it replaces the autoregressive process. In prior diffusion language models, such as Mercury, transformers are still utilized, but without causal masking. This allows the entire input to be processed simultaneously, resulting in a different approach to output generation. It’s likely that Gemini Diffusion also employs a transformer architecture, albeit in a novel way to achieve its remarkable speed and efficiency. The combination of speed and quality makes Gemini Diffusion a significant advancement in the field of language models. Its ability to generate complex, interactive content quickly suggests a wide range of applications, from rapid prototyping to real-time coding assistance. As more independent reviews and benchmarks become available, we will gain a clearer understanding of its capabilities and how it stacks up against other models in the market. For now, the initial impressions are highly promising.