Zhipu AI Introduces GLM-5.3 Flagship Model with 1M Context, Major Coding and Cybersecurity Gains
Zhipu AI’s official documentation has updated the GLM-5.3 model page, showing that this latest flagship model builds on the exact same base as GLM-5.2, achieving a capability leap through post-training at tens of times the scale of long-horizon task environments. The company says coding feel improves 50% versus the previous generation, and the model ranks first among open-source models on public benchmarks including Terminal Bench 3.0, while demonstrating emerging cybersecurity capabilities. The model API is coming soon, and GLM Coding Plan has fully rolled out GLM-5.3.
- GLM-5.3 builds on the same base as GLM-5.2 with post-training at tens of times the scale
- Coding feel improves 50% versus previous generation; Terminal-Bench 3.0 rises from 4.6 to 28.3
- 1M context window, up to 128K output tokens
- CyberGym score of 84.5%, cybersecurity capability approaching Mythos 5
What is GLM-5.3?
GLM-5.3 is Zhipu AI’s latest flagship base model, with text input and output, a 1M context window, and up to 128K output tokens. The model supports thinking mode, streaming output, Function Calling, context caching, structured output, and MCP tool invocation.
Breakthrough in coding and agent capabilities
According to official data, GLM-5.3 ranks among the leading open-source models across multiple mainstream benchmarks, with coding and agent capabilities comparable to Claude Fable 5. Terminal-Bench 3.0 rises from 4.6 to 28.3, DeepSWE v1.1 from 46.2 to 66.9, and Agents’ Last Exam from 23.8 to 28.5. It scores 1769 on GDPval-AA v2 covering 44 occupations, showing capability extending from coding toward professional task execution.
Beyond demos: delivering complex projects
GLM-5.3 can take on more complex, longer-term engineering goals, working continuously across tens of thousands of lines of code, hundreds of files, and multiple interdependent systems. For real-world tasks such as frontend development, bug fixing, and code refactoring, it can develop autonomously and iterate with minimal human intervention, driving projects to a deliverable state.
Post-training scaling in real expert workflows
GLM-5.3’s training is no longer limited to isolated coding problems, but extends to complete workflows from problem discovery and solution analysis to implementation, verification, and delivery. Some training tasks are equivalent to several consecutive days of work for a senior engineer, requiring the model to use real compute clusters, storage systems, internal documents, and codebases. Within a similar token budget, GLM-5.3 achieves a better balance between completion quality, execution speed, and cost.
Emerging stronger cybersecurity capabilities
As long-horizon task environments expand, GLM-5.3 shows better-than-expected capabilities in security tasks such as white-box code review, vulnerability discovery, and verification. Its CyberGym score reaches 84.5%, slightly above Mythos 5’s 83.8% and GPT-5.6 Sol’s 83.6%; ExploitBench rises from 24.4% to 54.4%. The company notes that current strengths concentrate on the front end of the exploit chain, with room for improvement in deeper exploitation and full offensive-defensive tasks.
Availability
The GLM-5.3 model API will launch soon, with complete call examples and a model experience center to follow. GLM Coding Plan has fully rolled out GLM-5.3.
Source: Zhipu AI Open Documentation
