What did Zhipu AI reveal on August 26, 2026, about the anonymous “Ox Alpha” model known as “Niu Lai,” including its identity as GLM 5.3 Flash, the model’s free anonymous testing on OpenRouter and OpenCode that attracted more than 50 trillion tokens in five days and later achieved

On August 26, Zhipu AI (Z.ai) disclosed that the anonymous “Ox Alpha”/“Niu Lai” model was its GLM-5.3-Flash: an open-weight, native-multimodal mixture-of-experts model tested anonymously before release. 146

What the reveal established

• The anonymous free deployment on OpenRouter and OpenCode drew extraordinary demand: reports put it above 50 trillion tokens within five days, and later above 60 trillion in six days; it led both platforms and reportedly peaked at more than twice comparable DeepSeek usage. 3
• Zhipu said the service ran entirely on a cluster of more than 100,000 Chinese-made chips. It claimed that, after software optimization, hardware efficiency and per-token cost were comparable with mainstream Nvidia GPUs. 36
• The exact chip suppliers were not disclosed. Huawei, Moore Threads, and Hygon were reported as possible suppliers, but this remains unverified speculation rather than a Zhipu confirmation. 40

Technical significance

• This was not proof that China has matched Nvidia hardware generally, nor that domestic-chip inference costs are universally equal to Nvidia’s. It was evidence that a large, real-world global inference workload could be served on a non-Nvidia domestic stack through substantial systems engineering. 36
• That matters for Nvidia’s CUDA advantage: it suggests that model-specific software stacks and inference engines can reduce dependency on CUDA for some production inference workloads. It does not erase CUDA’s broader ecosystem, tooling, developer-base, or training advantages. Insufficient evidence supports a stronger conclusion.
• Zhipu’s SGLang-based engine reportedly used W8A8 quantization, mixed INT8/FP8/BF16 cache quantization, intra-node tensor parallelism, and an Encode–Prefill–Decode (EPD) split into independently scalable pools. These techniques were intended to manage memory, bandwidth, communication, and scheduling constraints at a 1-million-token context length. 47
• Zhipu reportedly said these changes delivered roughly a threefold improvement in end-to-end service performance versus its initial baseline. The later SGLang documentation also reports higher throughput and materially greater FP8 cache capacity than a BF16 comparison configuration, though that is not an independent audit of Zhipu’s original threefold claim. 2

Model, performance, and price

• GLM-5.3-Flash has 320 billion total parameters and activates about 18 billion per token, with a 1,048,576-token context window. 57
• It reportedly received an Artificial Analysis Intelligence Index score of 57, placing it at the reported score of Claude Opus 4.8; this should be read as parity on that composite index, not as evidence of equal performance on every benchmark or real-world agent task. 1
• Its positioning was to compete with efficient frontier-class models such as DeepSeek V4 Flash and Pro while claiming substantially lower inference prices. Evidence supplied here is insufficient to verify exact benchmark-by-benchmark comparisons with both DeepSeek variants.
• Zhipu’s published API pricing was reported as $0.15 per million input tokens and $0.50 per million output tokens. 1
• Contemporary reporting characterized this as roughly one-tenth of GLM-5.3’s price and about one-fortieth of Claude Opus 4.8’s, but those ratios depend on the token direction, context tier, and pricing plan being compared. 1