Beijing AI Z.ai publishes a new generation of programming models GM-5.3. The company disclosed that the model had been placed online through the GMCing Plan subscription service and ZCode, and that the API and downloadable weights would be opened in stages after the completion of the security clearance.

Parameter size increased to 74.3 billion

According to the company, the GLM-5.3 was based on GLM-5.2, which was upgraded mainly by expanding the scale of training, rather than by re-engineering the bottom structure. The training focused on more environment, richer tasks and better calculator inputs, with emphasis on token efficiency.

According to Z.ai, the GLM-5.3 parameter is 743.0 billion. In the home Z.ai Code Bench test, the model scored 34.5 per cent under the maximum strength setting, up 23.4 per cent above the GLM-5.2. The single-task output token consumption is about 75,000, down from 96,000 in the previous generation.

The core list still lags behind the American closed source model.

In public comparison, GLM-5.3 has exceeded some open source or Chinese-like models, but in some core programming lists, the lead is still held by the United States closed source model.

GLM-5.3 scored 28.3 below the 33.7 and 34.6 of Fable 5 and GPT-5.6 Sol in a real Linux-like tool call test of Terminal Bench 3.0. GLM-5.3 scored 66.9 in the DeepSWE v1.1 test to fix GitHub's real problem, which was also slightly lower than the 67.5 and 69.7 of Kimi K3 and Fable 5.

Low prices and open weights attract developers

In addition to performance, price and open weight arrangements were the focus of the publication. According to Z.ai, GLM Coding Plan uses a credit system that halves the cost of non-peak-time transfers, and its API pricing is lower than some of the US front models.

The article mentions that the official GLM-5.2 price is token ($1.40 per million) and token 4.40 per million. By comparison, the GPT-5.3-Codex price is about $ 1.75 per million input token and $ 14 per million output. The GLM-5.3 model weights are expected to be available to the public in about two weeks, as indicated in the publication notes.

Additional information:Z.ai, based in Beijing, is included in the list of United States entities. This "open weight" refers to the version that will be released at a later date, with subscription services and tool portals on line and model weights not yet available for download.