Z.ai used GLM-5.3 agents to stand up inference on 100,000+ Chinese accelerators in under two weeks and tripled throughput
Z.ai documented building production inference for GLM-5.3-Flash on a cluster of more than 100,000 domestic AI accelerators, going from initial model adaptation to production in under two weeks with GLM-5.3-powered agents doing most of the porting work that normally takes an experienced infra team weeks. Optimizations (intra-node tensor parallelism for linear attention and the LM head, ReplaySSM, W8A8 quantization, mixed INT8/FP8/BF16 cache quantization, Layer Split, and an Encode-Prefill-Decode disaggregated architecture) delivered roughly 3x end-to-end throughput and per-token costs they describe as comparable to mainstream NVIDIA GPUs. Launched anonymously as 'Ox-Alpha' on OpenCode and OpenRouter, it became the most-used model on both, processing over 62 trillion tokens in six days.
Source
↳ Follow the thread