Chinese AI lab Z.AI just released GLM-5.2, and the numbers are hard to ignore. On FrontierSWE — a benchmark that tests whether AI can handle multi-hour engineering projects — it scored 74.4, barely trailing Claude Opus 4.8’s 75.1. It even beat GPT-5.5.
What makes this genuinely surprising? The whole thing was trained on Huawei Ascend chips. No Nvidia hardware anywhere in the pipeline.
Emad Mostaque from Stability AI estimated training costs at roughly $25 million — 80% of that spent on post-training. Compare that to what Western labs burn through for similar models, and it’s a different universe.
GLM-5.2 is a 744-billion-parameter mixture-of-experts model with a 1 million-token context window. That’s five times what GLM-5.1 offered. Whole-repo navigation and massive refactors become single-call operations instead of multi-step headaches.
API pricing sits at $1.40 per million input tokens and $4.40 for output. Claude Opus charges $5 and $25 respectively. A coding plan runs about $18/month and plugs into Claude Code, Cline, Kilo Code, and most agentic environments.
Can you run it locally? Technically, yes. Unsloth AI released 2-bit GGUF quantizations that shrink it from 1.51TB to 238GB while keeping roughly 82% accuracy. You’ll still need 256GB of RAM or a serious GPU setup with MoE offloading. Not exactly a weekend project, but doable with the right hardware.
Z.AI has been on the U.S. Entity List since early 2025. Their stock surged 90% over the past week. Whether that’s sustainable or hype catching momentum, the benchmark results speak for themselves.
