GLM-5.2 inference speed race hits 392 tokens per second

Jie Tang of Z.ai, the lab that makes GLM-5.2, is reacting to someone else serving his company's model faster than expected. His post quotes Dmytro Dzhulgakov of Fireworks, an inference provider that hosts open-weight models, who wrote that GLM-5.2 at 280 tokens per second was already considered impressive, then offered 318, adding that Fireworks still has room to go. Tang's reply calls the number crazy and says his own team has to work even harder.

The figure comes from an Artificial Analysis chart Dzhulgakov attached, which tracks how fast each hosting provider generates output for the same model. On June 25, 2026 it showed Fireworks at 318, ahead of Databricks at 288, Wafer at 271 and Baseten at 242, with several providers serving lower-precision FP8 versions further down. The chart shows many companies hosting the same model and competing on speed, which is why the model's maker is reacting to a third party's number rather than announcing its own. Artificial Analysis's provider comparison for GLM-5.2 is the ongoing version of that race. Neither post mentions 392 tokens per second; the number at stake here is 318.