We built the fastest API for GLM-5.2 (280 TPS) (www.baseten.co)

A new breakthrough in AI has emerged with the development of the world's fastest API for the GLM-5.2 model, achieving over 280 tokens per second (TPS). This impressive performance stems from a range of optimizations in the Baseten Inference Stack, including the use of a custom runtime engine, NVFP4 quantization on NVIDIA Blackwell GPUs, and innovative techniques like KV-aware routing and prefill-decode disaggregation through NVIDIA Dynamo. GLM-5.2 stands out not only for its speed but also for its cost-effectiveness, operating at 70-80% lower costs compared to competitors like GPT 5.5, which positions it as a game-changer for commercial AI applications. The GLM-5.2 model boasts 744 billion parameters and excels in complex tasks such as coding, while supporting a one-million-token context window. The enhancements in Multi-Token Prediction and disaggregation of prefill and decode processes lead to improved efficiency and responsiveness in production, allowing for optimized resource allocation during inference. These technological advancements not only elevate GLM-5.2's competitive edge but also signify a significant step forward in open AI models, enhancing the scope and viability for developers looking to deploy high-performance AI solutions.