版本历史记录
懒猫微服官方
GLM 5.3 Full 0.1.0
2026.11.15

GLM 5.3 Full: Original EXLR8 Full weights, dual Jetson AGX Thor, quad optical ports, and DFlash2 auxiliary decoding. Supports four concurrent sessions and a maximum request size of 12,224 tokens, providing native cumulative usage statistics. The model, runtime archive, and control plane are available from fixed-version dual sources on Hugging Face and ModelScope. The 30 TPS targets for 128, 2K, and 8K single streams have not yet been met—see the accompanying report for performance results.