Tencent Holdings has officially released the final version of its Hunyuan 3 (Hy3) foundation model, marking a significant upgrade in its artificial intelligence capabilities. The launch introduces enhanced autonomous-agent functionalities, which the Shenzhen-based technology giant has immediately integrated into its consumer AI assistant, Yuanbao. This release transitions Hy3 from its preview phase, which began in April, into a fully deployed enterprise and consumer tool, accompanied by a reduction in developer usage fees to encourage broader adoption across the Chinese tech ecosystem.
According to Caixin Global, the upgraded Hy3 model boasts a 90 percent task-completion rate across several of Tencent’s internal applications. This high success rate reflects improved post-training compute and higher-quality training data compared to the April preview version. The integration of these advanced agent capabilities into Yuanbao allows users to generate complex documents and execute multi-step workflows using natural-language prompts, positioning the assistant as a more capable digital coworker rather than a simple chatbot.
Architectural Advancements: 295 Billion Parameters, 21 Billion Active
The technical specifications of Hunyuan 3 reveal a massive scale and a sophisticated architecture designed for efficiency. Hy3 is a Mixture-of-Experts (MoE) model featuring 295 billion total parameters. However, its MoE design ensures that only 21 billion parameters are active during any given inference task, utilizing 192 experts with top-8 routing. This structure allows the model to handle complex reasoning, agentic workflows, and coding tasks while managing computational costs, making it economically viable to deploy at scale across consumer and enterprise products.
Tencent’s official press release highlighted that the model has been integrated into multiple core business units simultaneously, including WorkBuddy, CodeBuddy, Marvis, and WeChat, in addition to Yuanbao. The company also announced that it has open-sourced the Hunyuan 3.0 model under the Apache 2.0 license, making the full 295-billion-parameter architecture available to the broader developer community. The API is now available on Tencent Cloud’s TokenHub platform, and the company has slashed developer pricing to accelerate third-party adoption. This move aligns with a growing trend among Chinese tech giants to release open-source models to build developer ecosystems and challenge proprietary systems from OpenAI and Anthropic.
(Related: Alibaba’s Qwen3-6-Plus Breaks Token Records with Enhanced Agentic Capabilities)
Implications for Tencent’s AI Ecosystem
The official launch of Hy3 and its agentic features represents a critical step in Tencent’s broader AI strategy. By folding its AI lab directly into the Hunyuan team earlier this year, the company signaled a shift toward consolidating its research and development efforts to accelerate product deployment. The decision to offer the AI-agent feature for free within Yuanbao is a clear aggressive play to capture market share in the highly competitive domestic AI assistant space, challenging rivals like Baidu’s Ernie Bot and Alibaba’s Tongyi Qianwen.
The reduction in API pricing for developers utilizing Tencent Cloud TokenHub indicates a push to make Hy3 the foundational layer for third-party applications. By lowering barriers to entry and providing a highly capable, open-source MoE model, Tencent is seeking to establish Hunyuan as the default infrastructure for enterprise AI in China. The reported 90 percent success rate in internal office tasks serves as a strong proof of concept for enterprise clients looking to automate complex workflows.
The Hy3 launch also carries significance in the context of the global AI race. By open-sourcing a 295-billion-parameter model under a permissive Apache 2.0 license, Tencent is directly competing with Meta’s Llama series for developer mindshare outside China. The combination of strong benchmark performance, free agent functionality, and open weights positions Hy3 as a serious contender for enterprise AI deployments globally, not just in the domestic market. The simultaneous integration across WeChat, which serves over 1.3 billion monthly active users, means Hy3’s real-world deployment scale will be enormous from day one.
The pricing strategy accompanying the launch is equally aggressive. Tencent has set the Hy3 API rate at 1 yuan per million tokens for standard queries, significantly undercutting many competitors in the domestic market. This price point, combined with Yuanbao’s free-agent feature, reflects a deliberate strategy to prioritize ecosystem growth over near-term revenue. The company appears willing to absorb significant implementation costs in the short term to establish Hy3 as the default model for Chinese enterprise developers. If successful, this approach could generate substantial long-term revenue through Tencent Cloud infrastructure consumption, even if the model API itself remains a loss leader.
Looking at the competitive landscape, the Hy3 launch intensifies an already fierce battle among China’s technology giants for dominance in the AI assistant and enterprise AI market. Alibaba’s Qwen series, Baidu’s Ernie 4.5, and ByteDance’s Doubao are all vying for the same enterprise and developer customers. Tencent’s decision to deploy Hy3 simultaneously across WeChat, WorkBuddy, and its cloud platform represents a multi-front offensive that leverages its unique distribution advantages. No other Chinese company has WeChat’s reach, and the ability to expose a state-of-the-art model to over a billion users simultaneously is a structural advantage that cannot be easily replicated.
(Related: AI Agents Get ‘Digital ID Cards’ Under China’s First Connectivity Standard)