Z.AI Releases Open-Source GLM-5.2 Model for Long-Horizon Coding Tasks

Z.AI has introduced GLM-5.2, an open-source model released under an MIT license designed specifically for long-horizon tasks. [1] The model features a 1M-token context length and supports multiple thinking effort levels to balance performance with latency. [1] Its architecture incorporates a method called IndexShare, which reduces per-token FLOPs by 2.9 times at the 1 million context length by reusing the same indexer across every four sparse attention layers. [1] Additionally, updates to the model's MTP layer for speculative decoding increase the acceptance length by up to 20 percent. [1]

GLM-5.2's extended context training focused on coding-agent scenarios like automated research, complex debugging, and performance optimization. [1] On the FrontierSWE benchmark, the model scored 1 percent lower than Opus 4.8 while outperforming GPT-5.5 by 1 percent and Opus 4.7 by 11 percent. [1] Furthermore, GLM-5.2 outperformed both GPT-5.5 and Opus 4.7 on PostTrainBench, trailing only Opus 4.8 in the evaluation of post-training small models on a single H100 GPU. [1]

Z.AI's release positions a fully open-source model in direct competition with proprietary elite systems like Opus 4.8 and GPT-5.5. [1] By utilizing architectural optimizations like IndexShare to reduce FLOPs at a massive 1 million-token context length, the developers have prioritized sustained viability for agents over raw context size claims. [1] The benchmark results indicate that MIT-licensed open weights can now rival closed models on complex, multi-hour software engineering tasks such as kernel optimization and automated ML research. [1] What the sources don't address: How the model performs on generalized knowledge benchmarks outside of targeted, long-horizon software engineering domains.