Wild that the AI model from the Chinese company actually on the U.S. Entity List (Z ai's GLM 5.3) appears to given honest answers in censorship testing, whereas DeepSeek and, to a lesser extent Kimi, are committed to Chinese propaganda.
Testing CN open model censorship (on FeatherlessAPI to see if it's baked into model) on most sensitive CN queries I can think of, and wow DeepSeek was deceptive.
-GLM-5.3 Flash: Zero refusals or censorship
-Kimi K3 first gives propaganda, then when pushed snaps out of it,
