Z.ai Reveals GLM-5.3-Flash: 100,000-Chip Domestic Cluster Achieves 3x Performance Gain

On September 17, 2026, Z.ai published a long technical blog disclosing that the previously anonymously tested "Ox Alpha," namely the GLM-5.3-Flash model, now runs entirely on more than 100,000 domestic AI accelerators, with performance tripled compared with the initial baseline and per-token cost comparable to mainstream solutions.

Factual Reconstruction

GLM-5.3-Flash was released on August 26, 2026, and is the first native multimodal model in the GLM-5 series, using a hybrid sparse and linear attention architecture with 320B total parameters and 18B activated parameters. Before release, the model was tested anonymously under the name ox-alpha on OpenCode and OpenRouter, becoming the most-used model on both platforms within a week and processing over 62 trillion tokens in six days.

The blog notes that no one had previously operated a domestic accelerator cluster of this scale. Chip memory capacity and bandwidth were limited, the ecosystem was immature, and kernel support was incomplete, requiring engineers to guess undocumented driver behavior. The team built an in-house inference engine based on SGLang, while a GLM-5.3-driven Infra Agent completed extensive infrastructure optimization.

Mechanism Breakdown

Z.ai proposes a "dense feedback method" that integrates correctness tests, runtime logs, execution traces, microbenchmarks, and end-to-end metrics into a reproducible workflow, enabling the Agent to validate hypotheses locally. Feedback must meet three properties: being local, low-cost and timely to obtain, and supporting objective validation.

Specific cases include: comparing partitioned and non-partitioned kernel paths revealed that tl.dot in the Context Parallelism path of the KDA kernel used TF32 by default, causing precision issues; after the fix, it was merged upstream into Flash Linear Attention. In DeepEP v1.2.1, intranodedispatch and intranodecombine did not release the Python GIL; after the fix, the KV Transfer gap dropped from more than 20% to within 1%. The Agent distilled optimization skeletons from projects such as SGLang and applied division optimization and V-dimension tile merging to the KDA Decode kernel, achieving a 1.71x speedup.

Industry Impact

This practice shows that, under VRAM and ecosystem constraints, domestic accelerators can become production-ready within two weeks through in-house engines and AI-assisted optimization. The rapid iteration of the model from initial adaptation to production deployment relies on a combination of technologies including the Encode-Prefill-Decode decoupled architecture, W8A8 quantization, and mixed-precision caching.

Strategic Assessment

[Analysis] The above cases indicate that AI-assisted infrastructure construction may reduce reliance on mature ecosystems, but long-term stability at larger scale and under more complex workloads still needs to be validated; if similar methods are replicated on other domestic clusters, they will change the inference cost structure and supply chain landscape.