I worked on it for a more specialized task (query rewriting). It’s blazing fast.

A lot of inference code is set up for autoregressive decoding now. Diffusion is less mature. Not sure if Ollama or llama cpp support it.

How was the quality?