Tencent Launches Hy3 AI Model in Leadership Shift

Tencent Holdings Ltd. unveiled its first major AI model update under new leadership on April 24, launching an open-source foundation model called Hy3 Preview that prioritizes practical deployment over benchmark chasing.

The Caixin Global report details Hy3 Preview as a mixture-of-experts system with 295 billion total parameters (21 billion activated) capable of processing 256,000 tokens of context, now deployed across Tencent's ecosystem including its chatbot Yuanbao, which will switch from DeepSeek as its primary model.

Industry analysts note this strategic pivot—away from the "parameter race" dominating Chinese AI development—aims to close a user gap with domestic rivals like Alibaba and ByteDance, as the company shifts focus to embedding AI directly into products rather than chasing theoretical performance metrics (a problem that has plagued users for years, frankly). The model's 256,000-token context window means users can paste entire research papers into Yuanbao without hitting a token limit, a common frustration with competing models that max out at 128,000 tokens.

The release follows the recruitment of former OpenAI researcher Yao Shunyu to lead Tencent's AI division, a move highlighted by the South China Morning Post, which notes Hy3 represents Tencent's most powerful model to date, though it still lags behind U.S. leaders like OpenAI's GPT-4 in raw capability. Yao's team rebuilt training systems to improve stability and tackle business needs across Tencent's products, rather than chasing benchmark scores alone.

Unlike the trend toward trillion-parameter models, Hy3's 295 billion-parameter design uses a mixture-of-experts architecture that activates only 21 billion parameters per task, reducing computational costs while maintaining performance. This approach mirrors how users interact with the model: a smooth, responsive chat interface that doesn't require waiting for a 10-second load time (a common frustration with larger models). In internal testing, Hy3 turned data into visualizations and built a working web game from a simple prompt—tasks that typically require multiple steps in competing systems.

Tencent's decision to open-source Hy3 Preview—unlike Alibaba's closed model strategy—reflects a broader industry shift toward collaborative development, though the company has yet to clarify how it will monetize the model beyond its existing cloud and product integrations. In March, the company announced it would more than double AI spending to over $5 billion in 2026, signaling a commitment to long-term development rather than short-term wins. This investment comes as Chinese tech firms roll out new AI models more quickly and invest in software agents that can complete tasks with limited human input, a trend that has accelerated since the March reorganization of Tencent's research teams.

Whether users will actually pay for the underlying technology remains the real question, given the crowded market where free alternatives already dominate. Tencent's strategy of embedding AI directly into its products—like QQ and its coding tools—may prove more sustainable than competing on raw model size, but the company still faces the challenge of convincing developers to adopt Hy3 over established platforms like Meta's Llama series or Alibaba's Qwen.

Artūras Malašauskas is an AI Systems Integrator with 20+ years of production-grade web engineering experience. He has designed, shipped, and scaled enterprise Python/PHP systems for logistics, SaaS, and public-sector clients. For the past year, he has focused exclusively on AI integrations: deploying open-source LLMs, building generative media pipelines (image, audio, video), and engineering multi-agent workflows for real production environments. His standard: reproducibility, security, cost-efficient inference—no vaporware. He documents and evaluates emerging AI tooling, separating verified capabilities from marketing noise. Technical editor at: muza-ai.eu, ai-verslas.lt, ai-naujinos.lt Connect on LinkedIn