Thread · 5 stories · Aug 26 – Sep 6
An anonymous model led OpenRouter for six days, turned out to be Z.ai's GLM-5.3-Flash, and was claimed to have run entirely on Chinese chips.
Jump to timeline ↓For six days an unnamed model called ox-alpha topped OpenRouter with 23.2 trillion tokens. Z.ai then revealed it as GLM-5.3-Flash — a 320B MIT-licensed MoE pairing sparse and linear attention — and said it had been served entirely on 100,000+ domestic Chinese chips.
Alibaba's Qwen3.8-Flash-Next landed a day later on almost the same recipe, and Z.ai open-weighted GLM-5.3 two days after that. The episode became the reference point for how far Chinese labs had closed the price-performance gap.
