Zhipu open-sources GLM-5.3-Flash, Moore Threads achieves day-zero support

Zhipu AI recently open-sourced its GLM-5.3-Flash (320B-A18B) model, the first natively multimodal model in the GLM-5 series. With 320B total parameters and only 18B activated, it scored 57 on the Artificial Analysis Intelligence Index, matching Anthropic's Claude Opus 4.8. Previously tested anonymously as Ox-Alpha on OpenCode and OpenRouter, it set new call records on both platforms.

Moore Threads achieved day-zero optimization using its MTT S5000 AI training-inference card and MUSA software stack. The team customized operators for the KDA mechanism in the hybrid architecture using the MATE optimization engine, and integrated with SGLang-MUSA cache management to boost long-context inference efficiency.