Zhipu recently launched and open-sourced GLM-5.3-Flash (320B-A18B). Moore Threads officially announced today that, based on the MTT S5000 AI training-and-inference computing card and the MUSA software stack, it achieved Day-0 support, ultra-fast adaptation, and efficient operation of the model.
According to the introduction, as the first natively multimodal model in the GLM-5 series, GLM-5.3-Flash has 320B total parameters, with only 18B activated. It scored 57 on the globally authoritative Artificial Analysis Intelligence Index (AA Intelligence Index), placing it among the world's frontier models and matching Anthropic's most popular model, Claude Opus 4.8. In the hands-on evaluation on the self-developed Z.ai Code Bench, its coding performance was comparable to Claude Opus 4.8.
Previously, the model underwent anonymous public testing on OpenCode and OpenRouter under the codename Ox-Alpha, also known in the community as “Niu Lai.” Thanks to its outstanding reasoning speed and intelligent performance, it quickly became the most popular model of the week, setting new records for call volume on both platforms.
On the day GLM-5.3-Flash was released, Moore Threads' engineering team completed an architectural breakdown, core technology analysis, and review of its key operators. For the KDA mechanism used by linear attention in its hybrid architecture, the team used the MATE operator optimization engine to rapidly complete customized implementations and in-depth tuning of new operator forms, including state matrix updates and block-parallel scanning. The team also worked closely with the SGLang-MUSA cache management system to create seamless integration, unlocking KDA's efficiency advantages for long-context reasoning.
GLM-5.3-Flash open-source repository:
https://huggingface.co/zai-org/GLM-5.3-Flash