Chinese open-weight MoE models: Z.AI GLM-5.2, Moonshot Kimi K3, Baidu ERNIE 4.5, MiniMax H3

Z.AI GLM-5.2 (753B, MIT) beats GPT-5.5 on coding at 1/6 cost, captures 40% developer tokens on OpenRouter, doubles Kimi K3 on long-horizon browser use. Moonshot AI Kimi K3 (2.8T MoE) beats GPT-5.6 Sol and Claude on coding/browsing; weights live on HuggingFace with revenue-based license; White House accusations; hardware bottleneck (B300 scarcity); infrastructure open-sourced; architecture reveals LatentMoE, NoPE, attention residuals. Self-hosting reality check: most users cannot run it locally, but the model is a milestone. Baidu open-sourced ERNIE 4.5 (MoE, multimodal) under Apache 2.0 with full training/inference tooling (ERNIEKit, FastDeploy). MiniMax H3 open-weight video model (2K resolution, native stereo sound, cost <1/3 rivals) generates 15-second 2K clips with native stereo audio; open weights promised but not yet shipped. All four drive geopolitical debate on open-weight risks. Enterprise adoption by DoorDash, Coinbase, Cursor.