For most of 2025, outside observers described Tencent as the cautious giant of China's AI race — rich, patient, and oddly quiet while rivals rushed to ship consumer chatbots. Then its own numbers told a different story.
Behind the calm, Tencent had been reorganizing its research labs, open-sourcing a full family of models, and quietly building an app that now sits among China's most-used AI-native products.
Key takeaways
- Tencent's 2025 full-year capital expenditure hit ¥79.2 billion (≈ US$11 billion) and R&D reached ¥85.75 billion (≈ US$12 billion), both record highs.
- Its Yuanbao (元宝) AI-native app reported over 50 million daily active users during the Spring Festival period and 114 million monthly actives, per the company's 2025 results.
- The Hunyuan (混元) model family now spans full-modality open source, and Tencent is pushing an agentic (智能体) development platform for enterprises.
The flagship app
Tencent's consumer bet is Yuanbao (元宝), an AI-native assistant built on its self-developed Hunyuan (混元) large model and, since early 2025, also wired to DeepSeek. The company's full-year 2025 results put Yuanbao at more than 50 million daily active users during the Spring Festival window and 114 million monthly actives — placing it in the top tier of China's AI-native apps.
The strategy has shifted from scatter to focus. A year earlier, Tencent's stated view was that "every app has a chance" to become a super-app; by its annual staff meeting, chairman and CEO Pony Ma (马化腾) was arguing that models and AI products should be considered as one integrated system, with Yuanbao as the core vehicle for capturing the AI entry point.
Opening the model family
Underneath Yuanbao sits Hunyuan (混元), which Tencent describes as a full-modality open-source family covering text, image, video and 3D. One widely cited release, Hunyuan-A13B, uses a Mixture-of-Experts design with only 13 billion active parameters while aiming to match much larger models on performance.
The open-source move is partly defensive. Tencent's cloud business sells the tooling that lets enterprises build their own agents, and a credible, free model lowers the barrier to adopting that stack. At its 2025 cloud AI summit, Tencent upgraded its knowledge-engine product into the Tencent Cloud Agent Development Platform (腾讯云智能体开发平台), letting users have an agent break tasks into steps, plan a path, and call tools on its own.
Rewiring the lab
In December 2025 Tencent announced internal restructuring aimed squarely at the model race: new AI Infra and AI Data departments, plus a data-computing platform group. Yao Shunyu (姚顺雨), a researcher recruited from OpenAI, was named chief AI scientist overseeing AI Infra and the large-language-model group. The moves signal that Tencent is treating foundation-model work as a central, funded mandate rather than a side project.
President Martin Lau (刘炽平) has pushed back on the idea that Tencent is behind. Different models, he has argued, hold advantages in different scenarios, and Tencent's entrenched social and gaming businesses give it the cash flow to fund AI without panic.
An ecosystem, not a single product
Tencent's advantage is distribution. Beyond Yuanbao, its existing products are being rebuilt around models: QQ Browser's agent capabilities serve more than 130 million users, Sogou's input-method AI features have passed 100 million users, and the ima knowledge tool reports over 13 million monthly actives. On the enterprise side, Tencent says more than half of China's top media companies run agent workbenches on its cloud.
According to Li An, chief scientist at BrainNet (脑机网), "Tencent's real edge is not any single model but the habit layer — if an assistant lives inside the apps people already open fifty times a day, adoption is a distribution problem, not a capability problem."
Honest limitations
Figures for capex, R&D, and Yuanbao's user counts come from Tencent's 2025 full-year results as reported by Chinese state-affiliated outlets and Tencent's own channels; I have not independently audited them. The "top tier" and "first tier" rankings are the company's framing, not a verified third-party league table. Hunyuan benchmark and open-source claims originate from Tencent and were not re-tested here. The Yao Shunyu appointment and December 2025 reorganization are reported but I have not confirmed them against a primary SEC/HKEX filing. I have omitted speculative claims about future model versions.
The model under the app: Hunyuan Hy3
Yuanbao is only as good as the model behind it, and that model kept moving. In April 2026 Tencent released a Hunyuan Hy3 preview built on a Mixture-of-Experts design with 295 billion total parameters and 21 billion active, a 256K context window, and a SWE-bench score of 74.4; the full Hy3 followed on July 6, 2026, with Tencent pitching it as a step up in agentic (智能体) capability and deeper product integration. By then Yuanbao had already crossed 100 million monthly active users, up from the 114 million reported in the 2025 full-year results, and had wired agent features into Tencent's office and social surfaces. The point for investors is that the capex and the lab reorganization described above were not abstract: they produced a shipped, iterating model family inside a product with a habit-layer advantage most rivals cannot buy.
What readers can do now
- If you run a Chinese-facing product, test Yuanbao's APIs and the Hunyuan open weights before defaulting to a closed model — the ecosystem integrations may save integration work.
- Enterprises evaluating agent platforms should compare Tencent Cloud's Agent Development Platform against Alibaba and Baidu equivalents on your own RAG and tool-calling workloads.
- Watch Tencent's quarterly capex trend; a sustained US$11 billion-plus run rate would confirm that the "cautious giant" has fully committed.