Alibaba just mass-dropped a three hundred and ninety-seven billion parameter model that it says beats GPT-5.2, Claude, and Gemini on eighty percent of benchmarks. And it activates only seventeen billion parameters per token. That's like hiring one expert from a room of four hundred and getting the right answer almost every time.