Zhipu's GLM-5.2 Goes Open Source and Tops GPT-5.5: 744B Parameters, MIT License, Trained Entirely on Huawei Chips
A Chinese company just dropped a frontier model's weights on the internet — no region locks, no usage restrictions, MIT license, use it however you want — and it outscores GPT-5.5 on coding benchmarks.
Three key facts
Zhipu AI (Z.ai) released GLM-5.2, a 744B-parameter Mixture-of-Experts model with 40B active parameters per token, fully open-sourced under the MIT license. After an initial rollout to paying coding customers on June 13, the weights were released with no regional restrictions and no usage conditions. On SWE-bench Pro it scored 62.1, beating GPT-5.5's 58.6. On Terminal-Bench 2.1 it hit 81.0, within four points of Claude Opus 4.8's 85.0. On MCP-Atlas, which measures multi-tool agent workflows, it scored 77.0 — statistically level with Claude Opus 4.8's 77.8 and ahead of GPT-5.5's 75.3. It ranks first among open-source models on the Intelligence Index v4.1 with a score of 51.
The entire model was trained on Huawei Ascend chips using the MindSpore framework, with zero NVIDIA dependency. This means Zhipu's training pipeline is immune to further U.S. export controls. GLM-5.2 doubled the parameter count of its predecessor GLM-4.5, yet its inference cost is one-sixth that of GPT-5.5 — $0.83 per million input tokens and $5 per million output tokens.
Zhipu is the world's first publicly traded foundation model company. It IPO'd on the Hong Kong exchange on January 8, 2026, raising $558 million at a $7.1 billion valuation. Its stock hit an all-time high of HK$1,993 on May 29. In a Bloomberg interview on July 12, founder Tang Jie said genuine security comes from broad participation and open oversight, not technological barriers.
WangDou's Take
The benchmark numbers aren't the headline — beating GPT-5.5 by 3.5 points on SWE-bench Pro is honestly within noise range. What's impossible to ignore is the combination of MIT license plus Huawei chips. OpenAI's GPT-5.6 Sol charges $30 per million output tokens. GLM-5.2 charges $5, and you can download the weights and skip the API bill entirely. Training a model on U.S.-sanctioned hardware that matches American closed-source models, then giving it away for free to the entire world — that's not a technical statement, it's a geopolitical one. Tang Jie says "security comes from openness," which translates to: the harder you lock down, the more I open up, and let's see who blinks first. Whether GLM-5.2 can actually replace Claude and GPT in real-world engineering depends on deployment data over the next six months. Benchmarks are just the entry ticket.
Source: VentureBeat · Bloomberg · TechTimes