Z.AI has launched GLM-5.2, a powerful AI model competing closely with Claude Opus without using NVIDIA chips.
Overview
- GLM-5.2 scores 74.4 on the FrontierSWE benchmark, closely following Claude Opus 4.8 at 75.1 and surpassing GPT-5.5.
- This model is built exclusively on Huawei Ascend chips and does not rely on NVIDIA hardware.
- Unsloth AI's new 2-bit GGUF quantizations have reduced the model size from 1.51TB to 238GB, though it still demands 256GB of RAM or VRAM for operation.
On June 16, Z.ai unveiled its latest model, GLM-5.2, which showcases notable improvements over its predecessor, GLM 5.1. The Beijing-based company has been on the U.S. Entity List since January 2025, but recent restrictions on Anthropic Fable may have benefited Z.ai, as its stock surged by 90% to a record high following the launch.
Performance metrics for GLM-5.2 highlight its competitive edge. It achieved a score of 74.4 on the FrontierSWE benchmark, just shy of Claude Opus 4.8's score of 75.1 and outperforming GPT-5.5's 72.6. In SWE-bench Pro, which assesses the ability to autonomously resolve GitHub issues, GLM-5.2 scored 62.1, surpassing GPT-5.5's 58.6 and GLM-5.1's 58.4, solidifying its status as a leading open-source model in the Artificial Analysis Intelligence Index.
GLM-5.2 was developed entirely on Huawei Ascend chips, with Emad Mostaque, founder of Stability AI, estimating the training costs at around $25 million, primarily for post-training processes. Earlier reports indicated Z.ai's focus on training image models on Huawei's Ascend Atlas servers, further demonstrating its reliance on non-American technology.
The model boasts a 1 million-token context window, significantly larger than the 200K limit of GLM-5.1, and is made available under an MIT license, which avoids governmental access restrictions.
Target Audience and Pricing
This expanded context window allows developers to execute complex tasks, like whole-repo navigation and multi-file refactoring, in a single call. API pricing is set at $1.40 per million input tokens and $4.40 per million output tokens, offering a more affordable option compared to Claude Opus 4.8's pricing of $5 and $25, respectively. A Coding Plan is available for about $18 monthly, compatible with popular coding environments.
Local deployment remains a viable option, with Unsloth AI's 2-bit GGUF quantizations maintaining around 82% accuracy while compressing the model size. However, it still requires high-end machines equipped with 256GB of unified memory, making it a consideration for users with advanced setups.
In practical testing, GLM-5.2 demonstrated its capabilities by generating a game combining typing mechanics and shooting elements. While the user interface was not as polished as some competitors, the variety of scenarios and enemy types showcased its potential for generating diverse outputs.
For those interested, GLM-5.2 is available for testing on Z.ai with some limitations, and open-source weights can be accessed via HuggingFace under the MIT license, with quantized weights also available for download.
