Tencent released Hy3, a 295-billion-parameter Mixture-of-Experts model with just 21 billion active parameters, under the Apache 2.0 license. This move eliminates a major barrier that blocked enterprise adoption of Chinese open-weight models across Europe, the UK, and South Korea.
Previous Chinese model releases carried restrictive licenses that barred deployment in these regions entirely, forcing legal and compliance teams to reject otherwise competitive models. Hy3's Apache license removes this friction. The model competes directly with Alibaba's GLM-5.2 at half the size and outperforms it across most benchmarks, with coding tasks as the primary exception.
The efficient parameter structure matters here. With only 21 billion active parameters out of 295 total, Hy3 runs on modest hardware compared to its parameter count. This translates to lower inference costs and faster deployment cycles. For enterprises balancing performance against operational constraints, the efficiency gains are tangible.
Tencent's decision reflects a strategic shift in how Chinese AI labs approach the global market. License terms that exclude entire regions create friction for multinational enterprises and cloud providers. By adopting Apache 2.0, Tencent signals confidence that Hy3 can compete on merit rather than through legal gatekeeping. The comparison with GLM-5.2 becomes relevant precisely because legal teams no longer need to kill the evaluation early.
The coding weakness versus GLM-5.2 matters for specific use cases. Development teams prioritizing code generation should still benchmark carefully. But for general-purpose tasks spanning reasoning, instruction-following, and factual recall, Hy3's performance edge and license terms create a clear alternative to restricted models.
This represents a maturation moment for open-weight model strategy. As competition intensifies, licensing becomes a key differ