I didn't know about HRM until this article, but it looks incredibly impressive, even in its current state. A 27M parameter model that can be trained to perform difficult domain-specific tasks with better results than flagship reasoning models? That is compelling.