Tencent's Hy3 Model: A Leap Forward in AI Practicality

Tencent Unveils Next-Gen AI Model with Real-World Focus

Tencent's Huan Yuan research team has taken the wraps off Hy3preview, their newest open-source language model that prioritizes practical applications over theoretical benchmarks. This release represents the first major update since the company overhauled its AI infrastructure earlier this year.

Built for Actual Use Cases

The hybrid expert model combines fast and slow thinking approaches, packing 295 billion total parameters with 21 billion activated at any time. What sets it apart is its focus on solving genuine problems rather than chasing benchmark scores. "We're stepping away from easily 'scored' public rankings," explains the development team, instead evaluating performance through real-world testing across Tencent's product ecosystem.

Performance That Matters

Early results show Hy3preview making waves in several key areas:

• Complex Reasoning: The model aced China's prestigious Tsinghua University mathematics doctoral exam and national biology competitions
• Coding & Search: Shows 54% faster response times in developer tools while handling workflows up to 495 steps long
• Everyday Applications: From WeChat customer service to Peace Elite game NPCs, the model delivers more natural interactions

Affordable Intelligence

Perhaps most striking is the cost efficiency. Through architectural optimizations, Tencut has reduced operating expenses significantly. On their TokenHub platform, input processing costs just ¥1.2 per million tokens - making advanced AI capabilities accessible to more developers.

"This is just our first step," says Yao Shunyu, Tencent's Chief AI Scientist. "We're actively seeking community feedback to refine Hy3's official release later this year."

The model is already powering services across Tencent's portfolio including QQ Browser, Tencent Docs, and various AI assistant products. Third-party integrations with popular tools like OpenClaw further extend its versatility.

Key Points:

• Hybrid expert model with 295B parameters (21B active)
• Focuses on practical applications over benchmark scores
• Shows 40% better reasoning efficiency than previous versions
• Already deployed across Tencent's product ecosystem
• Priced affordably at ¥1.2 per million input tokens