CtrlK
Tencent AI refers to the artificial intelligence research, products, and services developed by Tencent Holdings Ltd., one of the world's largest technology companies. Headquartered in Shenzhen, China, Tencent has invested heavily in AI across its ecosystem of consumer applications and cloud services. The company's flagship AI effort is the Hunyuan family of foundation models, which spans large language models, image generators, video generation systems, and 3D asset creation tools. Tencent integrates these models into products used by over a billion people, including WeChat, QQ, Tencent Cloud, and the Yuanbao AI assistant.
Tencent was co-founded on November 11, 1998, by Ma Huateng (known globally as Pony Ma), Zhang Zhidong, Xu Chenye, Chen Yidan, and Zeng Liqing in Shenzhen, Guangdong Province, China. The company name "Tencent" derives from its Chinese name Tengxun (腾讯), which incorporates part of Ma Huateng's Chinese name and loosely translates to "galloping fast information." Ma Huateng graduated from Shenzhen University in 1993 with a degree in computer science and continues to serve as Chairman and CEO.
Tencent's first product was OICQ, an instant messaging client inspired by ICQ, which launched in February 1999. The product was later renamed QQ following a trademark dispute with AOL. From this foundation, Tencent expanded into gaming, social media, fintech, cloud computing, and digital entertainment, growing into a conglomerate with a market capitalization exceeding $500 billion as of early 2026 and approximately 105,000 employees.
Tencent's investment in artificial intelligence predates the generative AI boom by more than a decade. The company established multiple dedicated AI research organizations over the years.
In 2007, Tencent invested over RMB 100 million to establish the Tencent Research Institute, with campuses in Beijing, Shanghai, and Shenzhen. This was one of the company's earliest formal commitments to long-term technology research.
Tencent YouTu Lab was founded in 2012 and became one of Tencent's three core AI laboratories. Based in Shanghai and Hefei, YouTu Lab specializes in computer vision, including optical character recognition, image recognition, and image generation. The lab developed AI capabilities for internal products such as QQ and Qzone, as well as for external partners including WeBank, SF Express, and China Unicom.
Tencent AI Lab was formally established in April 2016 after preparations that began in 2015. The lab recruited more than 50 AI scientists and 200 AI application engineers. In March 2017, Tong Zhang, an expert in machine learning and former professor at Rutgers University, became the lab's director, and Dong Yu, a former Microsoft researcher specializing in speech recognition, was appointed vice director.
Tencent AI Lab focuses on four core research areas: computer vision, speech technology, natural language processing, and machine learning. In 2017, the lab opened its first U.S. research center in Seattle to advance work on speech recognition and NLP.
In March 2026, Tencent announced a major organizational restructuring, folding the AI Lab into the Hunyuan team to consolidate its AI efforts under a single, more focused unit.
Tencent officially unveiled the Hunyuan large language model on September 7, 2023, at the company's Global Digital Ecosystem Summit. The initial model featured over 100 billion parameters and was pre-trained on 2 trillion tokens. Dowson Tong, Tencent's Senior Executive Vice President, described the model as having "strong capabilities in Chinese language comprehension, content creation, logical reasoning, and task execution."
At launch, Hunyuan was made available to domestic enterprise users through APIs on Tencent Cloud, supporting a Model-as-a-Service (MaaS) offering. The consumer-facing chatbot remained in internal testing on an invite-only basis. Tencent also began embedding Hunyuan into several of its own products, including Tencent Cloud, Tencent Games, and WeChat.
On November 4, 2024, Tencent open-sourced Hunyuan-Large, also known as Hunyuan-MoE-A52B. This was a Transformer-based mixture of experts model with 389 billion total parameters and 52 billion activated parameters per forward pass. The model supports a context window of up to 256,000 tokens.
Hunyuan-Large was pre-trained on 7 trillion tokens, including 1.5 trillion tokens of synthetic data designed to improve performance in mathematics, coding, and multilingual tasks. Key technical innovations included a mixed expert routing strategy, key-value cache compression, and expert-specific learning rate scheduling.
On benchmarks, Hunyuan-Large outperformed LLaMA 3.1-70B across multiple evaluations and demonstrated comparable performance to the much larger LLaMA 3.1-405B model. On MMLU, Hunyuan-Large scored 88.4, compared to LLaMA 3.1-405B's 85.2. The model weights were released on Hugging Face in three variants: Hunyuan-A52B-Pretrain, Hunyuan-A52B-Instruct, and Hunyuan-A52B-Instruct-FP8.
On February 27, 2025, Tencent released Hunyuan-TurboS, the first ultra-large Hybrid-Transformer-Mamba MoE model deployed in a production environment. The model was scaled to 560 billion total parameters with 56 billion activated parameters, comprising 128 layers in an architecture that alternates between Attention, Mamba2, and FFN blocks.
The Mamba2 layers use a state-space model (SSM) architecture with linear sequence-length complexity, addressing the quadratic scaling challenges of pure Transformer models in long-context scenarios. Each FFN layer uses an MoE structure with 1 shared expert and 32 specialized experts, activating 1 shared plus 2 specialized experts per forward pass. The model was pre-trained on 16 trillion high-quality tokens and supports context lengths up to 256,000 tokens.
Hunyuan-TurboS achieved a top-7 ranking on the LMSYS Chatbot Arena with a score of 1356.
In mid-February 2025, Tencent launched the Hunyuan T1-Preview reasoning model on the Yuanbao app, followed by the official T1 release on March 21, 2025. Built on the TurboS base, T1 activates 52 billion parameters through dynamic expert routing, with specialist modules handling specific reasoning domains like mathematical logic and contextual analysis.
T1 achieved strong results on reasoning benchmarks: 87.2 on MMLU-PRO (second only to OpenAI o1), 69.3 on GPQA-Diamond for scientific reasoning, and 64.9 on LiveCodeBench for coding. The model generates at 60 to 80 tokens per second and charges 1 yuan (approximately $0.14) per million input tokens and 4 yuan per million output tokens.
Tencent released Hunyuan 2.0 on December 5, 2025. This MoE model features 406 billion total parameters with 32 billion active parameters and supports a 256K context window. The model ships in two variants: Think (optimized for complex reasoning) and Instruct (optimized for instruction following).
Key improvements include higher-quality pre-training data and a dual-stage reinforcement learning approach combining RLVR (Reinforcement Learning with Verifiable Rewards) and RLHF (Reinforcement Learning from Human Feedback). On the IMO-AnswerBench mathematics benchmark, Hunyuan 2.0 Think scored 73.4. The Tau2-Bench agent benchmark score jumped from 17.1 (in the previous generation) to 72.4.
As of March 2026, Tencent has announced plans to launch Hunyuan 3.0 in April 2026. The model is expected to contain approximately 30 billion parameters with a focus on in-context learning and agent usability. Tencent is also developing an advanced AI agent for WeChat alongside this release.
| Model | Release Date | Type | Total Parameters | Active Parameters | Key Features |
|---|---|---|---|---|---|
| Hunyuan (v1) | September 2023 | LLM | 100B+ | Dense | Chinese/English language, 2T training tokens |
| Hunyuan-Large (MoE-A52B) | November 2024 | LLM (MoE) | 389B | 52B | 7T training tokens, 256K context, open-source |
| HunyuanDiT | May 2024 | Image Generation | N/A | N/A | Bilingual diffusion transformer, multi-turn generation |
| HunyuanVideo | December 2024 | Video Generation | 13B | Dense | Text-to-video, open-source, 13B parameters |
| Hunyuan TurboS | February 2025 | LLM (Hybrid MoE) | 560B | 56B | Mamba-Transformer hybrid, 16T training tokens |
| Hunyuan T1 | March 2025 | Reasoning LLM | N/A | 52B | Deep reasoning, 60-80 tokens/sec generation |
| HunyuanImage 3.0 | September 2025 | Image Generation | 80B | 13B | Autoregressive MoE, 64 experts, bilingual |
| HunyuanVideo 1.5 | November 2025 | Video Generation | 8.3B | Dense | Lightweight, runs on consumer GPUs |
| Hunyuan 2.0 | December 2025 | LLM (MoE) | 406B | 32B | Think/Instruct variants, 256K context |
| Hunyuan3D 2.1 | 2025 | 3D Generation | N/A | N/A | Text/image-to-3D with PBR materials |
| Hunyuan 3.0 | April 2026 (planned) | LLM | ~30B | N/A | In-context learning, agent focus |
HunyuanDiT (Hunyuan Diffusion Transformer) is Tencent's text-to-image model, open-sourced in May 2024. The model uses a diffusion architecture parameterized with a transformer and combines a pre-trained bilingual CLIP encoder with a multilingual T5 encoder for text processing.
A distinguishing feature of HunyuanDiT is its strong bilingual support for both Chinese and English, with particular strength in rendering Chinese text within generated images. The model also supports multi-turn text-to-image generation, allowing users to refine outputs through iterative dialogue. An integrated multimodal large language model interprets multi-round conversations and generates updated prompts for image generation.
Subsequent versions included HunyuanDiT v1.2 (July 2024), which introduced a low-VRAM variant requiring only 6 GB of GPU memory, and integration with the Diffusers library in June 2024.
Tencent open-sourced HunyuanImage 3.0 on September 28, 2025. Unlike HunyuanDiT's diffusion-based approach, HunyuanImage 3.0 uses a unified autoregressive framework similar to the architecture powering large language models. The model features 80 billion total parameters with an MoE structure containing 64 experts, activating 13 billion parameters during inference.
HunyuanImage 3.0 supports precise Chinese and English text rendering, handles prompts exceeding 1,000 characters, and produces high-resolution outputs. An Instruct variant was released on January 26, 2026.
HunyuanVideo is Tencent's open-source text-to-video generation model. The inference code and model weights were first released on December 3, 2024. With over 13 billion parameters, HunyuanVideo was the largest open-source video generation model at its time of release.
The model uses a "Dual-stream to Single-stream" hybrid transformer design with a Full Attention mechanism for unified image and video generation. In professional human evaluations involving over 1,500 prompts assessed by 60 evaluators, HunyuanVideo outperformed Runway Gen-3, Luma 1.6, and several leading Chinese video generation models. It achieved a 64.5% motion quality score compared to Gen-3's 48.3%.
HunyuanVideo was integrated into the Diffusers library on December 17, 2024, and FP8 model weights were released on December 18, 2024, to reduce GPU memory requirements.
HunyuanVideo 1.5 was released in November 2025 as a more efficient successor. The model reduced the parameter count from 13 billion to 8.3 billion while maintaining top-tier video quality. A key technical innovation is Selective and Sliding Tile Attention (SSTA), which achieves a 1.87x speedup for 10-second 720p video generation compared to FlashAttention-3.
The 1.5 version is designed to run on consumer-grade GPUs. A step-distilled 480p image-to-video model released on December 5, 2025, generates videos in 8 to 12 steps, reducing end-to-end generation time by 75% on an RTX 4090. Training code and LoRA tuning scripts were also released for community fine-tuning.
The Hunyuan3D series enables text-to-3D and image-to-3D generation. Since the first open-source 3D models were released in November 2024, the series has surpassed 3 million community downloads on Hugging Face.
Hunyuan3D 2.0 introduced a two-component system: Hunyuan3D-DiT for large-scale shape generation and Hunyuan3D-Paint for texture synthesis. Hunyuan3D 2.1 became the first production-ready 3D asset generation model, with full model weights and training code released for community fine-tuning. It supports PBR (Physically Based Rendering) materials for production workflows.
HunyuanWorld 1.0, released on July 26, 2025, is described as the industry's first open-source 3D world generation model. It generates immersive, explorable 3D environments from text or image inputs and is compatible with standard CG pipelines.
Tencent launched the Yuanbao AI assistant app on May 30, 2024, as its consumer-facing chatbot powered by Hunyuan. Yuanbao provides AI-powered search (integrating WeChat Search and Sogou Search), document analysis (supporting PDF, Word, Excel, and other formats), content creation, and image editing features.
Yuanbao was subsequently integrated directly into WeChat, allowing users to add it as a "friend" within the app and interact with it without downloading a separate application. This integration provides over a billion WeChat users with direct access to Tencent's AI capabilities. By early 2026, the AI-powered search function within WeChat handled over 90% of all question-based queries on the platform.
Tencent also introduced Yuanbao Groups in January 2026 for AI-powered social interactions, and the Yuanbao chatbot supports summarizing WeChat official account articles, responding to images and files, and generating structured reports.
Tencent has embedded Hunyuan across a broad range of products and services.
WeChat users gained access to AI features through the Yuanbao integration, initially appearing as a "Red Packet Cover Assistant" before expanding to broader assistant capabilities. QQ's browser was upgraded to an "AI Browser" with QBot, an assistant built on both Hunyuan and DeepSeek models, offering AI-powered search, office productivity, learning, and writing functions.
Ima.copilot, an AI-native application from the QQ Browser team, offers integration with WeChat Official Accounts and instant messaging. Sogou Input Method, also owned by Tencent, has integrated Hunyuan for enhanced text prediction and composition.
Tencent Meeting now incorporates intelligent recording, an AI assistant, and multilingual translation, serving over 15 million monthly users. Tencent Docs similarly uses Hunyuan for document generation and editing assistance.
Tencent Cloud's AI Coding Assistant, powered by Hunyuan, is used by more than 50% of Tencent's internal programmers and reportedly enhances productivity by 40%.
Tencent Cloud offers AI capabilities to enterprise customers through several products and platforms.
Tencent Cloud provides access to Hunyuan foundation models via APIs, enabling businesses to build and fine-tune industry-specific models. The MaaS offering includes more than 50 solutions across 20 major industries. Tencent also released Hunyuan Turbo as a cost-optimized model service, which doubled training efficiency and reduced inference costs by 50% compared to earlier offerings.
Launched as part of Tencent Cloud's enterprise AI suite, Agent Development Platform 3.0 (ADP) enables organizations to build and integrate autonomous AI agents for customer service, marketing, inventory management, and research workflows.
Tencent unveiled Tairos, an embodied intelligence platform designed to serve as the "AI brain" for humanoid robots and other physical AI systems. Tairos provides perception, motion planning, and human-machine interaction capabilities for robotics developers.
Tencent Cloud introduced a full suite of computing, storage, and networking solutions branded as AI Infra, optimized for large-scale model training and inference workloads.
Tencent competes in one of the world's most crowded AI markets. As of April 2025, more than 3,700 registered generative AI tools operated in China. Tencent's primary competitors include:
| Company | Primary LLM | Consumer Chatbot | Key Differentiator |
|---|---|---|---|
| Baidu | ERNIE | Ernie Bot | Early mover in search-integrated AI |
| Alibaba | Qwen | Tongyi Qianwen | Aggressive open-source strategy, e-commerce integration |
| ByteDance | Doubao / Seed | Doubao | Short-video ecosystem (Douyin/TikTok) |
| DeepSeek | DeepSeek R1/V3 | DeepSeek Chat | Cost-efficient reasoning models |
| Tencent | Hunyuan | Yuanbao | WeChat ecosystem, social integration |
China's AI market has been characterized by aggressive price competition. In May 2024, ByteDance released models costing roughly 99% less than OpenAI's comparable offerings. Alibaba followed by cutting Qwen model prices by 97%. In December 2024, Alibaba reduced Qwen-VL pricing by 85%, and in April 2025, Baidu cut Ernie 4.5 Turbo pricing by 80%.
During the 2026 Chinese New Year period, major tech companies spent heavily to acquire AI chatbot users. Baidu allocated CNY 500 million (approximately $72 million) to promote Ernie Bot. Tencent doubled that figure, spending CNY 1 billion (approximately $145 million) on Yuanbao promotion. Alibaba outspent both with CNY 3 billion (approximately $434 million) for Qwen adoption.
Tencent's competitive advantage in AI stems from its dominant social ecosystem. WeChat serves over 1 billion monthly active users, giving Tencent unmatched distribution for AI features. Bloomberg Intelligence has noted that Tencent and Alibaba are positioned to lead China's AI software sector, outpacing Baidu as DeepSeek enables broader tech development. Tencent has focused on embedding AI into existing products rather than competing solely on model benchmarks, using AI to enhance advertising targeting, content recommendations, and enterprise productivity.
Tencent's total capital expenditure reached approximately 79 billion yuan ($11.49 billion) in 2025, up from 77 billion yuan in 2024. However, spending fell below internal targets due to constraints in acquiring advanced AI chips caused by U.S. export controls.
For 2026, Tencent plans to more than double its AI investment, with spending expected to exceed CNY 36 billion ($5.2 billion) specifically on AI. Goldman Sachs projects total Tencent capital expenditure will reach RMB 100 billion in 2026 and RMB 110 billion in 2027. In the fourth quarter of 2025 alone, investment in Yuanbao and Hunyuan large models reached 1.6 billion RMB, with annual spending totaling 1.8 billion RMB.
Tencent's strategy for 2026 centers on three priorities: upgrading Hunyuan foundation models (with Hunyuan 3.0 planned for April), deepening AI integration across WeChat and other consumer products through AI agents, and expanding Tencent Cloud's AI infrastructure globally with localized, sovereign computing resources.
Tencent has adopted a substantial open-source strategy for many of its Hunyuan models. The following models have been released with open weights on platforms like GitHub and Hugging Face:
These releases are generally made available under the Tencent Hunyuan Community License Agreement, which permits research, educational, and commercial use with attribution requirements.