Chinese AI lab Z.ai, formerly Zhipu, today released GLM-4.5 and GLM-4.5-Air as open-weight models under the MIT license, claiming the flagship ranks third globally across 12 benchmarks — and first among both domestic and open-source models.

GLM-4.5 packs 355 billion total parameters with 32 billion activated per token; the Air variant has 106 billion total with 12 billion active. Both use a hybrid reasoning design that toggles between thinking and non-thinking modes, and support a 128K context window. Z.ai says the architecture is agent-native, integrating reasoning, perception and action for multi-step tasks. The company claims the flagship can run on just eight Nvidia H20 chips, roughly half the size of DeepSeek’s comparable model.

The release lands in the same month that Moonshot open-sourced Kimi K2 and Alibaba shipped Qwen3-Coder, as Chinese labs push open-weight models into global competition. Neil Shah of Counterpoint Research says the wave is broadening strategic choices for global enterprises, calling GLM-4.5 open, unrestricted and able to be quantized for a range of hardware. Prabhu Ram of CyberMedia Research says the models are reshaping adoption at US and global companies by offering cost-effective alternatives to proprietary Western systems.

On the boards tonight, the argument is about the leaderboard: does “third globally” hold up across a broad benchmark suite, or is this the usual launch-day cherry-picking? Either way, another Chinese lab has put frontier-adjacent weights out under a permissive license — and that, more than the exact ranking, is what has people talking.