Podcast Episode
The model supports configurable reasoning modes - disabled, low, and high - allowing it to balance speed and depth depending on the task. Tencent says it delivers improvements in complex reasoning, instruction following, in-context learning, code generation, and agent capabilities.
In March, Tencent went further, dissolving its decade-old AI Lab and folding its researchers into the Hunyuan team under Yao's leadership. The consolidation was part of what chairman Pony Ma acknowledged as an effort to catch up after being slow to act on AI.
The release arrives amid an intensely competitive period for Chinese AI labs. Moonshot AI open-sourced its Kimi K2.6 model this week, whilst Zhipu AI's GLM-5.1 and DeepSeek's V3.2 have both made waves in recent months. For Tencent, HY3 Preview is a statement that its bet on a young scientist and a company-wide restructuring is beginning to bear fruit.
Tencent Open-Sources HY3 AI Model Built by Ex-OpenAI Scientist
April 23, 2026
0:00
2:06
Tencent has released and open-sourced HY3 Preview, a 295 billion parameter Mixture-of-Experts language model with only 21 billion active parameters and a 256,000 token context window. The model is the first major output from a sweeping AI reorganisation led by former OpenAI researcher Yao Shunyu, who joined Tencent as chief AI scientist in December 2025.
A Leaner, Smarter Model
Tencent has released and open-sourced HY3 Preview, its most capable language model to date. The model is a Mixture-of-Experts architecture with 295 billion total parameters but only 21 billion active at any given time, supporting a context window of up to 256,000 tokens. The architecture represents a deliberate shift from the company's previous Hunyuan 2.0, which used roughly 400 billion parameters. By activating only a fraction of its total capacity per query, HY3 is designed to deliver strong performance at lower computational cost, reflecting Tencent's focus on deploying AI efficiently across its product ecosystem.The model supports configurable reasoning modes - disabled, low, and high - allowing it to balance speed and depth depending on the task. Tencent says it delivers improvements in complex reasoning, instruction following, in-context learning, code generation, and agent capabilities.
The Yao Shunyu Effect
The release is the first tangible output of a restructuring that began when Yao Shunyu, a Tsinghua University graduate born in 1997 and formerly a researcher at OpenAI, joined Tencent in December 2025 as chief AI scientist. Yao reports directly to president Martin Lau Chi-ping and was given control of both large language model development and AI infrastructure.In March, Tencent went further, dissolving its decade-old AI Lab and folding its researchers into the Hunyuan team under Yao's leadership. The consolidation was part of what chairman Pony Ma acknowledged as an effort to catch up after being slow to act on AI.
Deployment and Competition
HY3 Preview is already integrated into Tencent products including its Yuanbao AI chatbot, which reached 50 million daily active users in February, and its CodeBuddy coding assistant, used by more than 12,000 internal engineers. The model is also available through third-party platforms including OpenRouter.The release arrives amid an intensely competitive period for Chinese AI labs. Moonshot AI open-sourced its Kimi K2.6 model this week, whilst Zhipu AI's GLM-5.1 and DeepSeek's V3.2 have both made waves in recent months. For Tencent, HY3 Preview is a statement that its bet on a young scientist and a company-wide restructuring is beginning to bear fruit.
Published April 23, 2026 at 10:12pm
