tencent/Hy3

Tencent has unveiled Hy3, a cutting-edge large language model that represents a significant step forward in open-source artificial intelligence development. Released under the Apache 2.0 license, Hy3 combines advanced architecture with practical accessibility, positioning itself as a competitive alternative to proprietary models while maintaining open-source principles that benefit the broader AI community.

Hy3 is a 295-billion-parameter Mixture-of-Experts (MoE) model featuring 21 billion active parameters and a 3.8-billion-parameter Mixture-of-Token-Experts (MTP) layer. This architecture allows the model to efficiently process information by activating only necessary parameters during inference, reducing computational overhead compared to dense models of equivalent scale. Following an initial preview launch in late April, Tencent's Hy Team incorporated feedback from over 50 products and implementations before the official release, demonstrating a commitment to real-world optimization and practical deployment scenarios.

• Cost-Effective Deployment: The MoE architecture enables organizations to run a 295B model with only 21B active parameters, dramatically reducing computational and energy requirements compared to traditional dense models.

• Open-Source Competition: Hy3's Apache 2.0 license removes licensing barriers, allowing researchers and developers to freely modify and deploy the model without restrictions.

• Improved Language Understanding: The innovative MTP layer architecture suggests enhanced performance in token-level decision-making and contextual understanding.

• Global AI Development: As a major contribution from a leading Chinese tech company, Hy3 demonstrates the accelerating pace of non-Western AI innovation and competition.

• Production-Ready Framework: Extensive feedback integration from 50+ products indicates the model is designed with practical deployment requirements in mind.

Hy3's release signals an important inflection point in the AI landscape where sophisticated, production-grade models are becoming accessible under permissive open-source licenses. This democratization of advanced AI capabilities enables smaller organizations and researchers to implement state-of-the-art language models without massive infrastructure investments. As competition intensifies in the global AI market, Tencent's commitment to open-source development challenges the dominance of closed, proprietary systems and accelerates innovation across the entire industry. The model's efficiency through MoE architecture also addresses growing concerns about AI's computational sustainability and environmental impact.