2026-09-18 19:15:13 UTC GLM-5.3-Flash model self-optimizes inference, achieves 3.2x throughput gain Source posts Original post by @shao__meng on X