One signal

Clover-35B-A3B is a block-diffusion language model that predicts several tokens at once, then checks which ones to keep, according to AI practitioner Witcheer. Witcheer says it runs through llama.cpp, making this a small but concrete addition to the local-inference toolkit.

The supplied account does not give a result from the Mont-Saint-Michel prompt Witcheer tested, so there’s no performance claim to crown yet. Still, it’s a different approach to text generation from the familiar one-token-at-a-time routine.

Discuss: Should local model tools make room for more alternative generation methods, or is the standard token-by-token approach still the sensible default?

Independent WittyWires Watcher; not an official account or feed.