Zhipu identifies Ox Alpha as GLM-5.3-Flash, run on Chinese GPUs
This digest was compiled by AI from multiple sources — links to the originals are below.
Zhipu (Z.ai) confirmed that the anonymous Ox Alpha model released on OpenRouter and OpenCode on August 20 is its GLM-5.3-Flash. The model served 100 trillion tokens per day for a week entirely on Chinese AI chips. The disclosure ends speculation that Ox Alpha was Microsoft's MAI 2.
Key Facts
- Zhipu (Z.ai) identified the anonymous Ox Alpha model as its GLM-5.3-Flash.
- Ox Alpha was released on OpenRouter and OpenCode on August 20 with a 1-million-token multimodal context window and a 131,072-token output limit.
- The model processed 100 trillion tokens per day for a week entirely free of cost.
- Zhipu stated that all traffic during the anonymous trial was served on Chinese AI chips.
- GLM-5.3-Flash reduces the KV cache size by a factor of 4.4x relative to GLM-5.3.
Model Identification
Zhipu (Z.ai) confirmed that the anonymous Ox Alpha model was its GLM-5.3-Flash. The model was released on OpenRouter and OpenCode on August 20. Ox Alpha offered a 1-million-token multimodal context window covering text, audio, and video. It also provided a 131,072-token output limit.
Chinese Chip Infrastructure
Zhipu stated that the model ran entirely on Chinese AI chips during its anonymous trial. The trial week capacity was around 100 trillion tokens per day. The disclosure suggests Huawei's Ascend-class GPUs likely powered the service.
Technical Architecture
GLM-5.3-Flash combines linear attention with sparse attention. Linear attention reads text sequentially and updates a compact KV cache. Sparse attention retains detailed information on a slower memory tier with an index for retrieval. The model uses an IndexPool that compresses key information into a dense digital index. It also employs Manifold-Constrained Hyper-Connections (mHC) to govern inter-layer communication.
2 sources
Zhipu identifies Ox Alpha as GLM-5.3-Flash, run on Chinese GPUs
