Interesting they get to rev this with the release of a new flash model. I'm speculating part of the distil pipeline includes the image gen stuff; that seems like internal tooling that will pay dividends over time, if true. New frontier model -> automatic new image model. Even if it's just incremental updates, it's good for both the product cadence and compounding improvements.
The confusion here is dense, 3.1 Flash Image is not 3.1 Flash.
The banana models (image) are a different than the mainline models, but the confusingly leverage the same naming scheme.
> the distil pipeline
I don't have inside info, but everything we've seen about gemini3.0 makes me think they aren't doing distillation for their models. They are likely training different arch/sizes in parallel. Gemini 3.0-flash was better than 3.0-pro on a bunch of tasks. That shouldn't happen with distillation. So my guess is that they are working in parallel, on different arches, and try out stuff on -flash first (since they're smaller and faster to train) and then apply the learnings to -pro training runs. (same thing kinda happened with 2.5-flash that got better upgrades than 2.5-pro at various points last year). Ofc I might be wrong, but that's my guess right now.
Interesting. Whatever they are doing it's a bit different than Anthropic and oAI, which is good for the consumer. I'm curious about their ML Ops internally; would be fascinating to learn more.