Z.ai launches GLM-5.2 trained on Huawei Ascend chips

Z.ai released GLM-5.2 on June 16. The model has 744 billion parameters, uses a mixture-of-experts architecture, and is published under an MIT license. Z.ai made the open-source weights and quantized versions available for download on Hugging Face and for limited testing on its platform.

GLM-5.2 provides a one-million-token context window. That is five times the 200,000-token limit of GLM-5.1. The model’s architecture and extended context are intended for workflows that need long documents, whole-repo navigation, multi-file refactors and long agentic pipelines.

On the FrontierSWE benchmark, which measures an AI agent’s ability to complete multi-hour technical projects, GLM-5.2 scored a dominance rate of 74.4. Claude Opus 4.8 scored 75.1 on the same test and GPT-5.5 scored 72.6. On SWE-bench Pro, which assesses autonomous resolution of real-world GitHub issues, GLM-5.2 scored 62.1; GPT-5.5 scored 58.6 and GLM-5.1 scored 58.4. On the prolonged engineering test SWE-Marathon, GLM-5.2 scored 13.0 compared with Opus 4.8’s 26.0. Aggregate rankings list GLM-5.2 as a leading open-source model on several quality indices.

The model was trained entirely on Huawei Ascend hardware; Z.ai reports no NVIDIA GPUs were used in the training pipeline. An industry estimate put total training costs at about $25 million, with roughly 80% of that attributed to post-training work. Z.ai has been on the U.S. Entity List since January 2025. The company’s stock rose about 90% in the week around the model’s release and recent U.S. restrictions on some AI services.

Compressed 2-bit GGUF quantizations reduce the raw model size from roughly 1.51 terabytes to about 238 gigabytes while retaining about 82% of the original accuracy, according to published files. Running the model locally typically requires about 256 GB of unified memory or a combination of system RAM and GPU VRAM with mixture-of-experts offloading. Cloud-hosted APIs are likely to be the more practical option for many users.

Z.ai set API pricing at $1.40 per million input tokens and $4.40 per million output tokens. The company offers a Coding Plan that starts at about $18 per month and integrates with several agentic development environments.

Early hands-on tests reported that GLM-5.2 produced more varied outputs on zero-shot creative prompts, generating diverse game states and shifting enemy behaviors in a prototype task. Reviewers noted that user-interface output quality trailed some competitors on single-shot polished outputs.

GLM-5.2 follows GLM-5.1 and joins other large models that provide extended context windows and open licensing. The open weights, quantized files and API access are available now.