Home / Technology / Tencent's Hy3: Open Source AI Unleashed Globally
Tencent's Hy3: Open Source AI Unleashed Globally
6 Jul
Summary
- Tencent released Hy3 under Apache 2.0, removing previous license restrictions.
- Hy3 offers strong performance with a smaller parameter footprint.
- The model emphasizes reliability metrics crucial for enterprise deployment.
Tencent has removed a significant barrier to enterprise AI adoption by releasing the full version of its Hy3 model under the permissive Apache 2.0 license. Previously, license terms had restricted access for many interested enterprises.
The 295-billion-parameter Mixture-of-Experts model now has no regional exclusions, making it globally available. This open release follows feedback from over 50 internal product teams who tested the April preview.
Hy3 showcases a reduced hallucination rate, down from 12.5% to 5.4%, and fewer commonsense errors, now at 12.7%. These improvements are attributed to refined data cleaning and training.
Economically, Hy3's 295B total parameters offer a competitive advantage over larger models like GLM-5.2. Its FP8 footprint is under 300GB, requiring less memory and compute power for production serving.
The model is designed to run on export-compliant hardware, such as Nvidia's H20-3e GPUs. This makes it legally accessible for Chinese companies and efficient for Western data centers.
Tencent's Hy3 is positioned as the premier open-weight choice for search and tool-heavy agent workloads and reliability-sensitive applications, offering frontier-adjacent capabilities without excessive infrastructure demands.
