[Zhipu Releases GLM-5.1 High-Speed API, Achieving 400 Tokens/s to Set a Global Speed Record] According to monitoring by Dongcha Beating, Zhipu has launched the GLM-5.1 high-speed API for select enterprise clients, with a model output speed reaching 400 tokens/s, setting a new global upper limit for end-to-end speed in official large model APIs. While fully retaining the capabilities of the original flagship model, this high-speed version is powered by a high-performance inference engine jointly developed by Zhipu and the TileRT team. This engine completely reconstructs the GPU operation scheduling mechanism, statically orchestrating the model at compile time into a persistent Engine Kernel that resides on the GPU. During single-card inference, computation, asynchronous I/O, and communication are all broken down into tile-level micro-tasks, with the kernel launched only once. Intermediate results between operators are directly passed through registers and shared caches, eliminating the latency bubbles caused by frequent kernel launches and GPU memory reads/writes in traditional inference. When scaled to multi-card setups, TileRT further extends the specialization parallelism approach to the entire 8-card NVL topology, transforming originally homogeneous GPU nodes into heterogeneous Workers assigned to different tasks. For the attention layer computation of GLM-5.1, the system assigns GPU 0 as a Sparse Index Worker, dedicated to sparse index construction and routing decisions, while GPUs 1 through 7 are assigned as MLA Workers, responsible for the compute-intensive stages. Communication is fully embedded within the tile-level task pipeline, achieving deep overlap between computation and cross-card communication. This high-speed version of the service is currently available to select enterprise clients on the Zhipu MaaS platform. In the future, this technology will further optimize FP8 inference and ultra-long context production environments, providing more deterministic performance support for latency-sensitive scenarios such as AI programming, real-time interaction, and real-time speech. [Original Link]