Nano Banana 2, Qwen3.5 medium models and a diffusion LLM

Four items this week, and I would rather give you four real ones than pad the list.

Nano Banana 2

Google released Nano Banana 2, officially Gemini 3.1 Flash Image, on 26 February. The idea is to bring the quality of the Pro image model to Flash-level speed and cost. The improvements Google highlights are accurate text inside images, keeping up to five characters and a set of objects consistent across a workflow, and output from small sizes up to 4K.

Text rendering is the headline for me. Image generators used to mangle any lettering, which ruled them out for anything that had to be read: signage, labelled diagrams, slides. If that is now reliable, a whole category of everyday business graphics moves from "hire a designer" to "ask the model and check the result."

I would still check every label. A model that spells correctly most of the time is more dangerous than one that always fails, because you stop looking.

Open models at a practical size

Alibaba followed the Qwen3.5 flagship with a medium series: 27B, 35B-A3B and 122B-A10B, under open licenses. The 35B-A3B in particular activates only about 3B parameters at a time, which is why people were running it on gaming laptops within days. This is the size class where a small company can host a capable model on its own box.

Inception launched Mercury 2, a language model that uses diffusion to generate text in parallel instead of one token at a time. Whether diffusion wins out is open, but latency is a real constraint for voice and live agents, and trying a different mechanism is worth watching.

A policy line

On 26 February Anthropic published a statement about its discussions with the US Department of War. In it the company described the uses it is not willing to support and why. I read it as a useful example of a vendor stating its terms in public, which is the information a buyer needs.

Looking across the week, there is a theme about fit. Nano Banana 2 is a fast, cheap image model good enough for most uses. The Qwen medium models bring strong open performance to a single machine. Mercury 2 tries to make latency disappear. None of these is the biggest model of the year, and all of them are more likely to end up in a real workflow than a flagship that costs ten times as much.

What I'd watch next: how Nano Banana 2's text accuracy holds up on non-Latin scripts, Arabic in particular.