“It produced a very lengthy chain of thought that was almost entirely incorrect, but had some interesting “ideas” in it”
Alibaba’s Qwen team released QvQ-72B-Preview, an openly licensed model that thinks out loud about images. It counted pelicans fine and produced a Game of Life reasoning chain that was almost entirely wrong. It runs on a 64GB MacBook Pro after a 38GB download. A long chain of thought is not a correct one, and QvQ prints both.