Z.ai releases GLM-5.1: the model that doesn't run out of steam

Z.ai has rolled out GLM-5.1, and the main difference between this model and previous generations can be described in one phrase: it doesn't run out of steam.

The usual story with AI coders is this: the model takes a task, applies familiar techniques, quickly delivers initial results, and hits a ceiling. Giving it more time is useless; it just spins its wheels. GLM-5.1 has broken this pattern. The developers purposefully trained the model to work on long-distance tasks where there are no clear metrics and it needs to decide for itself what to do next.

The results speak for themselves. On the SWE-Bench Pro benchmark, the new model outperformed GPT-5.4 and Claude Opus 4.6. But it's even more interesting to see how it handles tasks without a finish line.

In one test, it was given the skeleton of a vector database in Rust and asked to squeeze out maximum speed. The usual limit for such tasks is 50 iterations. GLM-5.1 worked on the project over 600 times, making thousands of tool calls. Instead of stopping, the model analyzed the logs itself, found bottlenecks, and changed the architectural strategy. The result — 21,500 QPS compared to 3,547 from the best previous result. A sixfold difference.

A similar picture emerged with the optimization of GPU kernels for ML tasks. The model continued to find ways to accelerate the code even after a thousand iterations, bringing the average speed boost to 3.6x. For comparison, standard torch.compile yields about 1.15x.

The most telling scenario had no numerical metrics at all. GLM-5.1 was asked to build a working Linux desktop environment in the browser from scratch. No templates, no hints, no mockups. For eight hours of continuous work, the model built the system: adding a file manager, terminal, text editor, system monitor, calculator, and even games. Moreover, it didn't just slap widgets together, but refined the styles and tied everything into a unified interface.

The essence of the update is not that the model became slightly smarter on short tasks. It learned not to give up when easy wins end, and to methodically refine the solution over hundreds of iterations. This makes it a real tool for complex projects where the result cannot be judged by the first draft.

The model is already available on HuggingFace and GitHub, and you can see it in action via the Z.ai API.