> meager hardware Qwen was made on a cluster about that size. And this is before anybody ever thought about optimizing the training process. (Currently it's j

> meager hardware Qwen was made on a cluster about that size. And this is before anybody ever thought about optimizing the training process. (Currently it's just pytorch analyst-as-coder slop, with extremely overprovisioned quantizations, etc.)