Tencent's Hy4 AI model surfaces in gray testing, specs undisclosed

Tencent has begun gray-testing Hy4 Preview, its next open-source Mixture-of-Experts model, inside the Yuanbao assistant app, Crypto Briefing reported. The release marks another step in China's race to build ever-larger foundation models.

Tencent hasn't disclosed Hy4's exact parameter count or architecture. Unofficial reports put the total near 770 billion parameters with 49 billion active at any one time, but Tencent has confirmed none of those figures. What the company has confirmed is the plan behind it: during its Q2 earnings report in mid-August, Tencent said a larger Hy4 model focused on stronger performance and multimodal capability was coming soon.

That plan builds on Hy3, a 295 billion-parameter MoE model with 21 billion active parameters and a 256K context window. Tencent open-sourced Hy3 under the Apache 2.0 license in April and gave it a formal release in July. Weekly usage jumped more than 68-fold once it moved past preview, and the model now runs Tencent's Yuanbao assistant, its CodeBuddy coding tool, and its WorkBuddy productivity app, with gains in reasoning, coding, and multi-step tasks.

• Hy3: 295B total / 21B active parameters, 256K context, Apache 2.0, open-sourced April 2026, formal release July 2026
• Hy3 weekly usage: up more than 68x after formal release
• Aug 12-13: Tencent confirms a larger Hy4 model is coming, in its Q2 earnings report
• Aug 20-21: Hy4 spotted in gray testing inside the Yuanbao app
• Hy4 specs: undisclosed by Tencent; unconfirmed reports put it near 770B total / 49B active parameters

Hy4 first turned up in Yuanbao's gray testing around August 20 and 21, labeled an "expert-level model" inside the app's selector. Other outlets have reported the model appeared to rank above both Hy3 and DeepSeek in that same list, though Tencent has released no benchmark data to confirm the ranking.

The Hy series is led by Tencent's chief AI scientist Yao Shunyu, part of a dual strategy that pairs in-house models with open-source alternatives from labs like DeepSeek. The pattern shows up across Chinese tech more broadly: Apple trained its own China model with Alibaba's help earlier this month, and demand for the chips behind all of it keeps climbing, with Nvidia's latest sales forecast calling for 70% growth.

No public benchmarks exist yet for Hy4, and Tencent hasn't said when gray testing ends. Given how fast Hy3 scaled after its own preview, that gap probably won't last long. The comparison matters beyond bragging rights: Hy3's open license let developers outside China build on it directly, and if Hy4 ships under similar terms, its real test will happen in benchmarks run by outsiders, not inside Tencent's own app.

Nothing here should be taken as financial advice — just information to consider.