GLM 5.1 High Speed claims to achieve 400 tokens/s without reducing model capability, with the breakthrough point being engineering optimization at the inference system level rather than model compression.