August 14, 2026 — Igeekphone: Chinese AI company Zhishu has officially released its new GLM-5.3 large language model, just hours after its founder teased that a new model was on the way.
According to Zhishu, GLM-5.3 delivers a major performance upgrade over the previous GLM-5.2, with improvements of around 50% across key evaluations. The company positions the new model as the strongest open-source model currently available, particularly highlighting its capabilities in software engineering, autonomous agents, terminal-based tasks, and other complex real-world workflows.
Zhishu also claims that GLM-5.3’s coding and agentic performance is approaching that of Claude Opus 5, although independent testing will be needed to verify the comparison.
GLM-5.3 Keeps the 740 Billion-Parameter Architecture
Despite the substantial performance improvement, GLM-5.3 does not reportedly increase the parameter count compared with GLM-5.2.
Both models are based on a 740-billion-parameter architecture, meaning Zhishu has not yet moved the GLM series into trillion-parameter territory.
Previous rumors had suggested that GLM-5.3 could potentially expand to more than one trillion parameters. That has not happened with the current release.
The larger parameter count could potentially be reserved for a future model such as GLM-5.5, although Zhishu has not officially confirmed such a roadmap.
Instead, GLM-5.3 appears to achieve its gains primarily through post-training and optimization, demonstrating that model performance can continue to improve substantially without simply increasing the underlying parameter count.
Zhishu Positions GLM-5.3 as a New Open-Source Leader
According to Zhishu, GLM-5.3 currently ranks at or near the top of several major open-source model evaluations.
The company says the model has made particularly strong progress in:
- Software engineering
- Complex coding tasks
- Terminal operations
- Autonomous agent workflows
- Cross-tool collaboration
- Long-running tasks
- Professional knowledge work
These capabilities are becoming increasingly important as AI models evolve from systems that simply generate text or code into agents capable of completing multi-step tasks with limited human intervention.
Rather than focusing exclusively on conversational quality, GLM-5.3 is being positioned as a model designed to actually perform work in real computing environments.
Major Gains on Terminal-Bench 3.0
One of the most dramatic improvements appears in Terminal-Bench 3.0, an evaluation designed to measure how effectively an AI model can complete complex tasks in real terminal environments.
According to Zhishu, GLM-5.3’s score increased from 4.6 with GLM-5.2 to 28.3.
That represents a substantial improvement and highlights the company’s emphasis on practical software-engineering capabilities.
Terminal-based tasks require models to go beyond generating a code snippet. They may need to inspect files, execute commands, troubleshoot errors, modify code, and iteratively work toward a desired result.
Strong performance in this category is therefore an important indicator of an AI system’s ability to operate as a practical software-development agent.
DeepSWE v1.1 Shows Stronger Software Engineering
GLM-5.3 also reportedly made significant progress on DeepSWE v1.1, an evaluation focused on longer-term software engineering and continuous code modification.
The model’s score reportedly increased from 46.2 to 66.9.
This improvement is particularly significant because real software development rarely consists of a single prompt followed by a finished answer.
Developers and AI coding agents often need to:
- Understand an existing codebase.
- Identify the source of a problem.
- Modify multiple files.
- Run tests.
- Analyze failures.
- Make additional changes.
- Continue iterating until the task is completed.
The higher DeepSWE score suggests that GLM-5.3 has improved its ability to handle this type of extended engineering workflow.
Better Performance on Agents’ Last Exam
GLM-5.3 also reportedly improved its score on Agents’ Last Exam, an evaluation covering multiple professional scenarios and emphasizing cross-tool collaboration and longer-running tasks.
Its score increased from 23.8 to 28.5.
Agentic benchmarks are increasingly important because they measure capabilities that are difficult to capture through conventional question-and-answer evaluations.
A successful agent must be able to understand a goal, break it into steps, use external tools, respond to intermediate results, and continue working until the objective is achieved.
GLM-5.3’s improvement in this area suggests that Zhishu is attempting to build a model capable of functioning as a more autonomous digital worker rather than simply an advanced chatbot.
1,769 Score on GDPval-AA v2
Another notable result comes from GDPval-AA v2, an evaluation covering 44 occupations and designed to assess performance on high-value professional knowledge work.
Zhishu reports that GLM-5.3 achieved a score of 1,769 on this benchmark.
The result is notable because it suggests that the model’s improvements are not limited to programming.
While software engineering appears to be one of GLM-5.3’s strongest areas, the GDPval-AA v2 result indicates that the model is also being optimized for broader professional tasks.
This is an important direction for the AI industry as developers increasingly seek general-purpose agents capable of handling tasks across multiple knowledge-work domains.
GLM-5.3 vs GLM-5.2
The reported benchmark results show substantial improvements compared with the previous generation:
| Benchmark | GLM-5.2 | GLM-5.3 | Improvement |
|---|---|---|---|
| Terminal-Bench 3.0 | 4.6 | 28.3 | +23.7 |
| DeepSWE v1.1 | 46.2 | 66.9 | +20.7 |
| Agents’ Last Exam | 23.8 | 28.5 | +4.7 |
| GDPval-AA v2 | — | 1,769 | New reported result |
The figures above are based on Zhishu’s reported evaluation results. Independent replication will be important for determining how GLM-5.3 compares with other leading models under standardized testing conditions.
Coding and Agentic Performance Are the Main Focus
Zhishu’s messaging around GLM-5.3 makes it clear that coding and agentic capabilities are central to the new model.
The company claims that its programming performance is now approaching leading proprietary models such as Claude Opus 5 and that its coding ability surpasses other domestic Chinese models.
If independent evaluations confirm these claims, GLM-5.3 could become a significant competitor in the increasingly crowded AI coding-agent market.
The open-source nature of the model would make the development particularly noteworthy.
Developers, research organizations, and enterprises could potentially deploy, customize, and integrate the model into their own systems without depending entirely on a proprietary API.
Performance Gains Without a Larger Model
Perhaps the most interesting technical aspect of GLM-5.3 is that its improvements reportedly come without increasing the model’s parameter count.
Both GLM-5.2 and GLM-5.3 are said to contain 740 billion parameters.
This suggests that Zhishu has focused heavily on post-training techniques, data quality, reasoning optimization, tool-use training, and agent-specific capabilities.
The development reflects a broader trend in AI research: simply making models larger is no longer the only path toward improved performance.
Better training strategies can potentially extract considerably more capability from an existing architecture.
Could GLM-5.5 Move Into the Trillion-Parameter Range?
The absence of a trillion-parameter architecture in GLM-5.3 has also fueled speculation about future GLM models.
Some observers may interpret the current 740-billion-parameter architecture as a sign that Zhishu is deliberately reserving a major hardware and architecture upgrade for a future generation.
However, there is currently no official confirmation that GLM-5.5 will exceed one trillion parameters.
For now, GLM-5.3 demonstrates that Zhishu can achieve substantial performance gains without taking that step.
Why GLM-5.3 Matters for the Global AI Market
The release of GLM-5.3 comes as competition between open-source and proprietary AI models continues to intensify.
Companies around the world are developing increasingly capable models, while open-source projects are attempting to close the performance gap with commercial systems.
A model that combines strong coding performance, autonomous tool use, and professional knowledge-work capabilities could have applications across:
- Software development
- IT operations
- Automated testing
- Research
- Data analysis
- Business process automation
- Customer support
- Enterprise productivity
- AI agent platforms
The ability to execute tasks rather than simply generate responses is likely to become one of the most important competitive dimensions for the next generation of AI models.
Final Thoughts
The release of GLM-5.3 marks a significant step in Zhishu’s effort to compete at the highest level of the global AI industry.
Despite retaining the same 740-billion-parameter scale as GLM-5.2, the new model reportedly delivers major improvements in software engineering, terminal operations, agentic workflows, and professional knowledge work.
Its reported gains on Terminal-Bench 3.0, DeepSWE v1.1, and Agents’ Last Exam are particularly notable, while the 1,769 GDPval-AA v2 score points toward broader professional capabilities beyond programming.
Zhishu’s claim that GLM-5.3 is now the strongest open-source model and that its coding and agent performance is approaching leading proprietary systems is ambitious. Independent benchmarks and real-world testing will ultimately determine how it compares with the world’s top AI models.
Nevertheless, GLM-5.3 demonstrates an increasingly important trend: AI progress is no longer solely about making models larger. Through improved post-training and agent-oriented optimization, Zhishu is attempting to extract significantly greater practical capability from the same 740-billion-parameter foundation.
If these reported results hold up under independent evaluation, GLM-5.3 could become one of the most important open-source AI releases of 2026.