Fastest Among the Top: Zhipu AI, Are You 'Spraying' Code?

Zhipu AI has released the GLM-5.1-highspeed API, a high-speed version of its flagship model, claiming a blazing 400 tokens per second — the fastest among top-tier AI models. The name "highspeed" is no exaggeration: in a demonstration, thousands of lines of code appear on screen almost instantaneously, as if being "sprayed" out. This contrasts sharply with traditional AI coding, which often generates text line by line at a pace reminiscent of CPU rendering.

Speed alone might raise concerns about quality, but the results speak for themselves. The generated code runs effectively, handling complex tasks with ease. For instance, a single prompt requesting a webpage with a breathing nebula, audio-reactive particles, and adjustable parameters is executed in one burst after a brief thinking period. The model processes front-end structure, Canvas animation, state management, visual parameters, and interaction logic simultaneously, producing a polished product.

The speed advantage becomes even more apparent during iterative design. When users issue vague commands like "make the ripple faster" or "soften the particle dispersion," GLM-5.1-highspeed understands intent and adjusts the code in real time. This creates a feel akin to sitting beside a designer and tweaking parameters together, a level of interactivity previously hindered by latency.

Pushing further, the model demonstrates the ability to act as a game director. In a 2D game scenario with obstacles, enemies, and weather, the model can change the game world dynamically based on natural language commands like "snow" or "explosion." Such real-time world state modification becomes feasible only with sub-second response times, though engineering challenges like stability, safety, and cost remain.

On the content creation front, the API processes a 10,000-character article in just 10 seconds, generating multiple outputs such as poster headlines, short-video scripts, product slogans, and a complete WeChat article. This speed shifts human effort from waiting to judging, enabling more iterative collaboration.

The underlying technology is not merely a smaller model running faster. Zhipu's GLM team, in collaboration with the TileRT team, optimized the inference engine, scheduling system, and infrastructure. The inference engine rewrites core paths for the GLM-5.1 architecture, while the scheduling system uses dynamic batching, request merging, and KV cache scheduling to reduce tail latency. TileRT further pushes optimization to the tile level, leveraging compile-time static scheduling and persistent GPU kernels to minimize scheduling, data movement, and synchronization overhead. Think of it as pre-planned assembly lines instead of workers waiting for a foreman's instructions.

However, speed is not the sole metric. Production-grade APIs must also excel in model quality, stability, cost, context length, tool-calling reliability, concurrency, and error rates. The 400 tokens/s figure needs continuous validation across diverse tasks and conditions. Nevertheless, this release signals a clear trend: competition among domestic AI model APIs is shifting from mere correctness to delivering both speed and reliability. GLM-5.1-highspeed demonstrates that when flagship model capability meets high-speed inference, the AI agent experience transforms: less waiting, more feedback, and smoother task progression. In the age of coding, speed is the ultimate thrill.