Half the size, no legal fine print: Tencent’s Hy3 challenges GLM-5.2

Tencent has released the full version of its Hy3 language model under the permissive Apache 2.0 license, removing regional restrictions that previously excluded the European Union, the United Kingdom and South Korea. Sam Witteveen reports for VentureBeat that the license reversal, more than the model’s capabilities, drove the strongest reaction from the open-model community.

Hy3 is a 295-billion-parameter Mixture-of-Experts model with 21 billion active parameters, built on feedback from over 50 product teams following an April preview. In Tencent’s own blind test against 270 experts, Hy3 outscored the older GLM-5.1. But rival GLM-5.2, released by Zhipu AI in June, still leads on agentic coding benchmarks like SWE-bench Verified, according to Tencent’s own appendix.

Reliability over raw benchmarks

Tencent emphasizes production metrics instead: the company says hallucination rates fell from 12.5% to 5.4% compared to the preview, and multi-turn error rates dropped from 17.4% to 7.9%. It also reports consistent performance across different agent frameworks, an advantage for enterprises with mixed toolchains.

• Hy3: 295B parameters, 21B active, under 300GB in FP8
• GLM-5.2: roughly 744B parameters, 40B active, about 744GB in FP8
• Hy3 runs on Nvidia’s export-compliant H20-3e chips

That smaller footprint means Hy3 can run on far less hardware than GLM-5.2, which typically requires an 8x H200 node. Combined with the unrestricted license, Witteveen writes, Hy3 becomes an attractive option for reliability-focused, tool-heavy workloads, even as GLM-5.2 keeps its lead in coding. Independent verification of Tencent’s benchmark claims is still pending.