XNews

Zhipu GLM-5.3 Self-Optimized Inference on 100K+ Domestic Accelerators

Zhipu disclosed that GLM-5.3-Flash runs on over 100,000 domestic AI accelerators. From first run to production in under 2 weeks, an AI agent driven by GLM-5.3 performed all optimizations, achieving 3.2x throughput. Key optimizations include ReplaySSM, tensor parallelism, mixed INT8/FP8/BF16 cache, and encode-prefill-decode separation. The agent found and fixed three real engineering bugs.

Source posts