What's this about?
Google published the technical report for DiffusionGemma, a diffusion-based variant of its open Gemma model family.
What exactly is a diffusion language model?
Standard language models generate left to right, one token at a time, each conditioned on everything before it. Diffusion models work the way image generators do: start from noise and iteratively refine the whole output at once. For text this is a genuinely different bet, with two attractive properties. Generation can be parallel rather than strictly sequential, which changes the latency profile — and the model can revise earlier parts of its output as later parts take shape, where an autoregressive model is committed to every token it has already emitted. The open question is whether it matches autoregressive quality at scale, which is why a technical report from a major lab matters more than another demo.
Why is this worth noticing?
- It's a real architectural alternative from a lab with the resources to evaluate it properly, in a field where nearly everything shipped is autoregressive.
- The latency argument connects to this month's speed thread — Ultrafast at 750 tokens/second, the Cerebras partnership — where sequential generation is the fundamental constraint being worked around rather than removed.