Chinese AI company Z.ai has released GLM-5.2, an open-source model designed for coding tasks that run across longer workflows.
The model is available under an MIT license and supports a one million-token context window. Z.ai said it is aimed at work involving large codebases, multi-step engineering tasks, and extended reasoning sessions.
Built for longer coding tasks
The one million-token context window is relevant because Z.ai is positioning GLM-5.2 for project-scale engineering work, not only longer prompts. In software development, that means a model can process larger codebases and retain more task history within a single workflow.
The model can also take in related documentation and tool outputs. That matters for coding agents, which often need to move between code, commands, and test results during a task.
Z.ai said GLM-5.2 was trained for long-horizon coding-agent scenarios. The company listed use cases including large-scale implementation, automated research, performance optimization, and complex debugging.
GLM-5.2 follows GLM-5.1, which was also positioned around coding-agent workflows. Z.ai said the new model improves long-horizon task performance and adds multiple thinking-effort levels.
The release includes High and Max effort modes. Z.ai said these settings allow users to choose between faster responses and more compute-intensive processing for harder tasks.
Benchmark gains over GLM-5.1
The benchmark chart published by Z.ai gives the release a more specific developer angle. GLM-5.2 is not being presented only as a larger-context model, but as one aimed at software engineering workflows involving code inspection, tools, and command-line tasks.
The company has published benchmark results comparing GLM-5.2 with GLM-5.1 and several other models. On SWE-bench Pro, Z.ai lists GLM-5.2 at 62.1, compared with 58.4 for GLM-5.1.
On Terminal-Bench 2.1, Z.ai said GLM-5.2 scored 81.0, compared with 62.0 for GLM-5.1. The company also listed a best reported harness result of 82.7 for GLM-5.2.
The larger gain was on Terminal-Bench 2.1, which tests command-line software engineering tasks. SWE-bench Pro showed a smaller improvement, based on Z.ai’s published figures.
The same chart also helps keep the comparison measured. Z.ai said GLM-5.2’s Terminal-Bench 2.1 score was close to Claude Opus 4.8’s 85.0, but still below it.
The clearest comparison is with GLM-5.1, where Z.ai reports gains across several coding benchmarks.
Lowering long-context costs
Z.ai also said GLM-5.2 includes architecture changes aimed at reducing compute costs during long-context use. One of these is IndexShare, which reuses the same indexer across groups of sparse attention layers.
According to Z.ai, IndexShare reduces per-token FLOPs by 2.9 times at a one million-token context length. That claim is relevant to coding workflows where memory and compute needs rise as models process larger repositories.
The model also includes changes to its multi-token prediction layer. Z.ai said those changes increased the acceptance length for speculative decoding by up to 20%.
These changes are relevant because long-context coding agents can become more expensive to run as task histories grow. Repeated tool outputs can also add to the amount of context a model has to process.
Self-hosting options
Hugging Face documentation lists several ways to run GLM-5.2. The model can be used with Transformers, vLLM, SGLang, Docker Model Runner, and KTransformers, among other supported tools.
Documentation also lists support for deployment on Ascend NPU platforms. Supported frameworks include vLLM-Ascend, xLLM, and SGLang.
The open-source release gives developers the option to run the model on infrastructure they control. This differs from closed models, where access is generally provided through a hosted service or managed platform.
Self-hosting can give enterprise developers more control over deployment and data handling. It also shifts infrastructure management and tuning responsibilities to the user.
Developer testing still needed
GLM-5.2 has drawn attention from several technology executives and developers on social media. Vercel CEO Guillermo Rauch wrote on X that he was impressed by the model’s coding performance.
Matt Velloso, a former executive at Meta, Google DeepMind, and Microsoft, wrote that he had used GLM-5.2 for a full day and described it as an open model that met his bar for daily use.
Those comments point to early developer interest, but the model’s broader performance will depend on independent testing and real-world deployment.
GLM-5.2 is part of a broader group of Chinese open models that have gained attention from developers and AI companies. DeepSeek’s R1 previously drew attention for its reasoning performance and low-cost release.
GLM-5.2 is more narrowly focused on coding-agent tasks and long-context software engineering. Z.ai’s published results show gains over GLM-5.1 across several listed coding benchmarks.
The next test is whether those results translate into consistent performance in production coding workflows.
(Photo by Fotis Fotopoulos)
See also: Three AI coding agents launched in 72 hours, just changed what developers pay for intelligence
Want to dive deeper into the tools and frameworks shaping modern development? Check out the AI & Big Data Expo, taking place in Amsterdam, California, and London. Explore cutting-edge sessions on machine learning, data pipelines, and next-gen AI applications. The event is part of TechEx and co-located with other leading technology events. Click here for more information.
Developer Tech News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.