Moore Threads Announces Day-0 Support for Zhipu AI's GLM-5.3-Flash Model
Moore Threads has announced immediate "Day-0" support and optimization for Zhipu AI's newly released GLM-5.3-Flash model on its MTT S5000 GPU and MUSA software stack. The engineering team quickly adapted the model's hybrid architecture, specifically optimizing its KDA (Kimi Delta Attention) linear attention mechanism. This rapid adaptation demonstrates the agility of China's domestic GPU ecosystem in supporting state-of-the-art local AI models immediately upon release. It highlights the growing viability of domestic hardware-software stacks like MUSA as alternatives to mainstream platforms like NVIDIA's CUDA. To optimize the KDA linear attention mechanism, Moore Threads utilized its MATE operator optimization engine to implement custom operators like state matrix updates and block parallel scanning. These optimizations were integrated with the SGLang-MUSA cache management system to improve efficiency during long-context inference.
## BACKGROUND
Moore Threads is a Chinese GPU developer that created MUSA, a proprietary software stack designed as an alternative to NVIDIA's CUDA. The GLM-5.3-Flash model utilizes Kimi Delta Attention (KDA), a hardware-efficient linear attention mechanism that uses fine-grained, channel-wise gating to maintain a fixed-size memory state regardless of sequence length.