Zhipu AI Confirms GLM-5.3-Flash Trained and Served on 100,000 Domestic Chinese Chips

Zhipu AI Confirms GLM-5.3-Flash Trained and Served on 100,000 Domestic Chinese Chips

Zhipu AI disclosed that GLM-5.3-Flash, the open-weight model it released Wednesday, had already run an unannounced trial across OpenRouter and OpenCode under the alias Ox Alpha, processing 62 trillion tokens on code generation and complex task routing before the company claimed the checkpoint as its own. The entire test deployment ran on a cluster of 100,000 domestic Chinese chips rather than foreign-made accelerators, and shares tied to the Beijing-based startup rose after the disclosure.

The detail that matters here is operational, not the model itself. Zhipu is one of the more active labs tracked in the AI Market Watch index, with 35 items logged about the company in the last 90 days versus 17 in the prior 90 (per the AI Market Watch index; coverage is name-matched over pipeline-ingested sources, not a census). AMW covered GLM-5.3-Flash's initial launch on August 27; this story adds the missing piece — proof that a 100,000-chip domestic cluster can carry both public inference traffic and training-scale throughput, the exact question Chinese labs have faced since export restrictions cut off reliable access to foreign GPUs.

For builders, the open-weight release lets enterprises self-host GLM-5.3-Flash and skip recurring per-token API fees, though independent benchmarks against commercial rivals are still outstanding. For investors and infrastructure watchers, Zhipu's cluster becomes a reference point other Chinese model developers will cite when arguing their roadmaps aren't compute-constrained by trade policy — a claim whose credibility now depends on benchmarks nobody outside Zhipu has yet run.

#ZhipuAI #GLM53Flash #OpenWeightAI #ChinaAIChips #ExportControls #AIInfrastructure