Tencent has unveiled Hy4 Preview, an open-source Mixture-of-Experts (MoE) model. The release marks a significant escalation in China’s intensifying AI arms race, where the country’s biggest tech companies are racing to build ever-larger foundation models.
For context, Tencent’s previous Hy3 model clocked in at 295 billion total parameters. Think of parameters as the knobs and dials that a model can tune during training: more parameters generally means a model can learn more nuanced patterns, though how efficiently those parameters are used matters just as much.
From Hy3 to Hy4: a rapid evolution
The Hy4 Preview builds on what was already a strong foundation. Tencent’s Hy3 model, a 295B total parameter MoE architecture with 21B active parameters and a 256K context window, was open-sourced under the Apache 2.0 license in April 2026 before receiving its formal release in July.
Weekly usage of Hy3 rose more than 68-fold after it transitioned from preview to formal release.
Hy3 became the engine behind a growing suite of Tencent products, powering features across Yuanbao (Tencent’s AI assistant app), CodeBuddy (its coding assistant), and WorkBuddy (its enterprise productivity tool). The model delivered notable improvements in reasoning, coding, and handling complex multi-step tasks.
Gray testing and the road to full release
During its Q2 earnings report around August 12-13, 2026, Tencent disclosed plans to launch a larger-parameter Hy4 model imminently, with a focus on enhanced performance and multimodal capabilities.
By August 20-21, initial gray testing of the Hy4 model was spotted in the Tencent Yuanbao app, where it was positioned as an “expert-level model.” No official specifications, including the exact parameter count and architecture details, have been disclosed.
The Hy series is led by Tencent’s chief AI scientist Yao Shunyu. Tencent has adopted a dual strategy: developing proprietary models like the Hy series while simultaneously leveraging top open-source alternatives, including models from DeepSeek.