Z.ai announced the launch of GLM-5.2, its latest flagship AI model built for long-horizon tasks and complex agentic workflows.
The model is designed to support project-scale engineering work, longer coding-agent trajectories, and multi-step tasks that require sustained reasoning across large amounts of context. Z.ai said GLM-5.2 represents a major improvement over GLM-5.1 and is built to make long-context AI more reliable for real engineering use cases.
A key feature of GLM-5.2 is its 1 million-token context window, which is intended to help the model handle large codebases, extended workflows, and project-level software engineering tasks. The company said the goal is not only to support more tokens, but to maintain stable performance when working through long and complex tasks.
GLM-5.2 is positioned for agentic coding, long-horizon planning, software engineering automation, and workflows that require AI systems to understand requirements, follow engineering standards, generate code, test outputs, and complete broader development tasks over time.
The model also introduces flexible thinking effort levels, allowing users to balance performance, latency, and compute needs. Z.ai said this enables developers to choose stronger reasoning for difficult tasks while using lower-latency settings when speed and efficiency matter more.
Z.ai also highlighted improvements to the model’s architecture. GLM-5.2 uses IndexShare, a technique designed to reduce per-token compute requirements at long context lengths. The company also improved the model’s multi-token prediction layer for speculative decoding, which can increase acceptance length and support more efficient inference.
On coding benchmarks, Z.ai said GLM-5.2 improved significantly over GLM-5.1. The company reported stronger performance on Terminal-Bench 2.1 and SWE-bench Pro, while also positioning GLM-5.2 as one of the strongest open-source models for long-horizon coding tasks.
Z.ai said GLM-5.2 performs strongly on long-horizon benchmarks such as FrontierSWE, PostTrainBench, and SWE-Marathon. These benchmarks evaluate whether models and agents can complete larger engineering projects, improve smaller models through post-training, and handle extended software engineering tasks such as compiler development, kernel optimization, and production-grade service creation.
The company is releasing GLM-5.2 as an open model under an MIT open-source license, which Z.ai said is intended to provide broad technical access without regional restrictions.
GLM-5.2 is available through Z.ai and can also be used through developer and model platforms that support the release. The launch positions Z.ai to compete more directly in agentic coding, open-source AI, and long-context enterprise development workflows.
With GLM-5.2, Z.ai is focused on moving AI development from shorter coding assistance toward longer, more autonomous engineering execution. The company said the model is designed for the next phase of agentic AI, where systems need to understand large projects, sustain context, execute multi-step workflows, and deliver more complete software outcomes.
