web3: Z.ai Releases GLM-5.3 Programming Model Targeting Open Source Weight Market

CoinWorld reports:

Beijing AI company Z.ai has launched its next-generation programming model GLM-5.3. The company disclosed that this model has already been made available through the GLM Coding Plan subscription service and ZCode, with the API and downloadable weights to be gradually opened after completing security reviews.

Parameter Scale Increased to 74.3 Billion

The company stated that GLM-5.3 is built upon GLM-5.2, primarily upgrading by expanding the post-training scale rather than redesigning the underlying architecture. The training focuses on more environments, richer tasks, and higher computational power investment, while emphasizing token usage efficiency.

According to Z.ai, GLM-5.3 has a parameter count of 74.3 billion. In its own Z.ai Code Bench tests, the model scored 34.5% under maximum intensity settings, surpassing GLM-5.2's score of 23.4%. The single-task output token consumption is approximately 75,000, lower than the previous generation's 96,000.

Core Rankings Still Lag Behind U.S. Closed-Source Models

From public comparisons, GLM-5.3 has surpassed some open-source or similar Chinese models, but in some core programming rankings, the leading positions are still occupied by U.S. closed-source models.

In tool invocation tests in real Linux environments like Terminal Bench 3.0, GLM-5.3 scored 28.3, lower than Fable 5's 33.7 and GPT-5.6 Sol's 34.6. In the DeepSWE v1.1 test, which fixes real issues on GitHub, GLM-5.3 scored 66.9, slightly lower than Kimi K3's 67.5 and Fable 5's 69.7.

Low Prices and Open Weights Attract Developers

In addition to performance, pricing and open weight arrangements are also key points of this release. Z.ai introduced that the GLM Coding Plan adopts a points system, with costs for calls during non-peak hours halved, and its API pricing is lower than some leading U.S. models.

The article mentions that the official price for GLM-5.2 is $1.40 per million input tokens and $4.40 per million output tokens. In comparison, GPT-5.3-Codex is priced at about $1.75 per million input tokens and $14 per million output tokens. According to the release notes, the model weights for GLM-5.3 are expected to be made public in about two weeks.

Additional Information: Z.ai is headquartered in Beijing and has been placed on the U.S. Entity List. The "open weights" refer to versions that will be released later; the current offerings are subscription services and tool access, with model weights not yet available for download.