Overview
- Alibaba on Aug. 26 published Qwen3.8‑Flash and released Flash‑Next weights as a Qwen4 prototype, pitching a 125 billion‑parameter MoE model that activates about 6 billion parameters per token and supports up to a 1 million‑token context window.
- Z.ai on Aug. 26 confirmed the anonymous 'Ox Alpha' trial was GLM‑5.3‑Flash, published MIT‑licensed weights for the 320 billion‑parameter hybrid model that activates about 18 billion parameters per token, and said the trial ran at very large scale on domestically produced AI chips.
- Both vendors emphasize low inference costs and aggressive API pricing that make self‑hosting and on‑premises deployment more practical for enterprises, with Alibaba and Z.ai offering token prices well below many Western frontier APIs.
- The demonstrations matter because they pair open weights with production‑scale serving on Chinese accelerators, showing a path to chip sovereignty that reduces dependence on Nvidia for high‑volume inference.
- The rapid releases accelerate an open‑weight playbook that broadens developer access but intensifies trade‑offs for companies and regulators over data residency, licensing, export controls and the safety risks of widely distributable, agent‑capable models.