ReplaySSM Speeds GLM and Kimi Inference Up to 2.1x in Production

A new computational method called ReplaySSM has launched across various AI models and deployment frameworks, increasing standard decoding speed by up to 1.47x and speculative decoding by up to 2.1x. The technique is active in GLM-5.3-Flash and Kimi-K3, and is integrated into vLLM, TensorRT-LLM, SGLang, and FlashInfer. Developed in collaboration with Nvidia, the method caches State Space Model inputs instead of the state to separate output computation from state updates during decoding.