Overview
- Z.ai disclosed Wednesday that the anonymous Ox Alpha endpoint was its GLM‑5.3‑Flash and posted the model weights on Hugging Face under an MIT license for public download and reuse.
- The company describes GLM‑5.3‑Flash as a natively multimodal reasoning model with a one‑million‑token context window, a hybrid sparse/linear attention design and architecture tuned for very low per‑token inference cost.
- Independent fingerprinting groups found an exact tokenizer match to GLM‑5.x and measured a sharply concentrated censorship pattern that blocks or alters outputs on seven China‑domestic sensitive topics, including prompts about Xi Jinping.
- Z.ai says it served the anonymous OpenRouter/OpenCode trial at extreme scale using only domestically produced Chinese accelerators, a claim that observers and reporters have not independently verified while markets pushed Z.ai’s shares higher.
- The release accelerates developer uptake of an open‑weight frontier model and highlights policy and operational challenges—companies face data‑routing and export‑control risks, and the episode underscores China’s push for chip self‑sufficiency and the governance questions around widely available powerful models.