AIWiki
Malaysia
Back to all articles
Companies & Toolstencenthunyuanyuanbao

Tencent AI

4 min readUpdated October 2026
Tencent AI
Type
Artificial intelligence division of a technology group
Parent
Tencent Holdings Ltd (Shenzhen)
Flagship model
Hunyuan (Hy3, Hy4 Preview)
Consumer assistant
Yuanbao
Enterprise products
Tencent Cloud AI, WorkBuddy, CodeBuddy, ima
AI Lab established
2016
Tencent AI is the artificial intelligence organisation of Tencent Holdings, the Chinese technology conglomerate behind WeChat, QQ and one of the world's largest gaming companies. Its central research product is the Hunyuan family of foundation models, which powers Tencent's consumer chatbot Yuanbao, its enterprise assistants WorkBuddy and CodeBuddy, Tencent Cloud's model-as-a-service platform, and AI features embedded across WeChat, QQ, Tencent Docs and Sogou.

History

Tencent established its AI Lab in 2016 and published hundred-billion-parameter natural-language models by 2021, but its modern foundation-model effort dates from the launch of the original Hunyuan large language model on 6 September 2023, among the first batch of Chinese models filed with regulators under the country's generative-AI interim measures.[1]

Rather than treating Hunyuan only as a closed product, Tencent adopted a mixed strategy: a consumer assistant, Yuanbao, launched in May 2024, alongside a steady stream of open-weight releases — the Hunyuan text-to-image model (2024), the Hunyuan Large mixture-of-experts model (389 billion total parameters, 52 billion active), a 13-billion-parameter text-to-video model, the Hunyuan-A13B hybrid-reasoning model (2025), and the Hunyuan-MT-7B translation model, which topped Hugging Face's global model popularity chart in September 2025.[2]

In late 2025 Tencent appointed former OpenAI researcher Vinces Yao as Chief AI Scientist, restructured its model organisation, and shipped Hunyuan 2.0 in December 2025. During 2026 the release cadence accelerated to roughly a major version every two months after an infrastructure rebuild: Hy3, a 295-billion-parameter MoE model with 21 billion active parameters, launched in July 2026, followed by the open-sourced Hy4 Preview in August 2026 — 770 billion total parameters, 49 billion active, context beyond one million tokens.[3]

Key Concepts and Technology

Hunyuan models are built on the transformer architecture and increasingly on mixture-of-experts (MoE) designs, where a sparse subset of specialised parameter blocks is activated per token, giving large total capacity at tractable inference cost.[4] Tencent publishes technical reports alongside major open releases and has open-sourced specialised components — OCR, 3D and world generation, translation, video and avatar models — making Hunyuan one of the broadest open model portfolios from a Chinese lab.

Equally important is integration depth. By the end of 2025 Hunyuan was deployed in more than 900 internal Tencent applications, and from 2026 Tencent repositioned it as the substrate of an internal agent fleet rather than a single user-facing chatbot: WorkBuddy (an AI office assistant, personally championed by founder Pony Ma), CodeBuddy for software development, and the ima knowledge workspace all draw on Hunyuan while also routing to other models.[1]

Applications and Impact

Tencent's AI reach spans gaming (NPC and content generation tooling), advertising and content recommendation, fintech, Tencent Cloud's enterprise API customers, and productivity software used across Chinese offices. Its large capital expenditure on chips and inference capacity during 2026 made Tencent one of the heaviest AI spenders among global internet companies, and its open-weight releases are widely mirrored on aggregation platforms such as OpenRouter and Hugging Face.[3]

>See Also

🇲🇾Malaysian Context

🇲🇾 Tencent's AI relevance to Malaysia is largely commercial and ecosystem-level. Tencent is a long-standing major shareholder of Sea Limited, operator of Shopee and Garena, both of which serve Malaysian users, and WeChat's mini-program and payment ecosystem is used by Malaysian sellers trading with Chinese consumers — channels where Hunyuan-based translation, customer service and content tools are increasingly offered.[5]

Malaysian enterprises evaluating Chinese foundation models face the same governance questions as elsewhere: model provenance, cross-border data flows and compliance with the Personal Data Protection Act 2010 when personal data is processed on offshore clouds. These considerations sit alongside Malaysia's own sovereign-model policy, which backs domestic efforts such as Ilmu through the National AI Office while leaving businesses free to use foreign APIs.[6] For developers, the practical angle is straightforward: Hunyuan's open weights can be self-hosted locally, and Malaysian AI practitioners commonly benchmark Hy-series models alongside DeepSeek and Qwen on Malaysian-language tasks.

References

  1. ↑Baidu Baike. (2026). Tencent HY (Hunyuan). https://baike.baidu.com/en/item/Tencent%20HY/1450766
  2. ↑Tencent Hunyuan. (2024). Hunyuan-Large: An Open-Source MoE Model with 52 Billion Activated Parameters. arXiv. https://arxiv.org/abs/2411.02265
  3. ↑LinkedIn. (2026). Tencent unveils Hunyuan Hy3 open-source large language model. https://www.linkedin.com/posts/mervinpraison_tencent-has-unveiled-hunyuan-hy3-its-latest-activity-7478248158017744-shTt
  4. ↑Hugging Face. Transformers documentation: Mixture of Experts (MoE) explained. https://huggingface.co/docs/transformers/en/moe
  5. ↑Sea Limited. Official website. https://www.sea.com/
  6. ↑National AI Office Malaysia. Official website. https://ai.gov.my/
Back to all articles