On April 23, 2026, Tencent released and open-sourced the Hunyuan Hy3 Preview language model. Within two weeks, it topped both Token call volume and market share on OpenRouter, the world largest API aggregation platform. Call volume exceeded 10 times that of its predecessor Hy2. This is the first model after the Hunyuan team comprehensive reconstruction of pre-training and reinforcement learning infrastructure, led by former OpenAI researcher Yao Shunyu.
Hy3 Preview is a MoE (Mixture of Experts) model that integrates fast and slow thinking, with 295B total parameters, 21B activated parameters, and a maximum context length of 256K tokens. Inference efficiency has improved by 40%, with native support for programming tool calls and complex Agent tasks.
Tencent has shifted away from benchmark chasing, building over 50 customized test benchmarks internally to evaluate real-world industrial applicability. From architecture upgrade to official release, Hy3 Preview took less than three months.
Tencent stock rose over 3% on the release day. Goldman Sachs issued a report calling Hy3 Preview a key milestone, maintaining a buy rating with a target price of HKD 700.
