Ox Alpha Revealed as Zhipu's GLM-5.3-Flash, Running on Chinese Chips All Along

eyishazyer · x · 2026-08-27

The mystery model Ox Alpha has been confirmed by Bloomberg as Zhipu's GLM-5.3-Flash. The poster spent six days piecing it together from tokenizer leaks and weird error codes before Bloomberg confirmed it outright, with weights dropped that evening.

The specs are aggressive: 320B total parameters with only 18B active, 1M token context, native text/image/video multimodality, MIT licensed, priced at $0.15 in / $0.50 out per million tokens. The arguably bigger headline: the whole stealth week ran on Chinese AI chips, not Nvidia. Benchmarks show real gains over GLM-5.2 and genuine competitiveness with Opus 4.8, GPT-5.6 Terra, and Gemini 3.7 Flash — though exact deltas are contested (a viral 10-task DeepSWE sample claimed 80%). The thread also previews a stacked release run: Qwen's Paloma, a possible Opus 5.1, Astra, and Grok 4.7.

Original post →

More from Models

Models channel →