Tencent has kept a low profile in the large model race for years. But at this year’s World Artificial Intelligence Conference (WAIC), the company signaled a clear shift: it is no longer staying quiet about artificial intelligence applications.
On July 26, WAIC opened in Shanghai, and Tencent leaned in with a full-throated pitch. Its message centered on AI agents serving as digital companions for China’s 1.4 billion people. Through these agents, Tencent aims to build a network that integrates AI into users’ daily routines.
At the event, the company introduced a suite of new offerings, including proprietary foundation models and AI agents, as part of what it described as a full-stack AI platform built to meet a “family bucket” of needs.
“AI is evolving from short-term to long-term memory,” said Wu Yunsheng, vice president of Tencent Cloud and head of Tencent Youtu Lab. He noted that for a long time, large models struggled to retain extended context, making complex tasks more difficult.
Tencent’s focus on AI agents reflects a broader shift in the technology’s trajectory. Wu pointed to the rise of multi-agent collaboration, where AI moves beyond basic Q&A interactions toward richer, multimodal exchanges across video, imagery, and audio. He said the next phase will involve agents specializing in different tasks and working together to solve more intricate problems.
When Tencent Cloud introduced the Hunyuan model family in 2023, it emphasized industry-specific solutions, offering more than 50 tailored models for sectors such as finance, public services, and telecommunications. That narrative has now expanded. The company’s ambitions have grown to include multimodality and embodied intelligence.
This year’s WAIC marked Tencent’s most significant presentation to date on embodied AI. The company’s Robotics X lab and Futian Laboratory jointly launched Tairos, a modular, plug-and-play platform that integrates foundation models, development tools, and data services. Marketed as the first of its kind in China, the platform targets the robotics industry, aiming to close a key software gap for hardware and application developers.
A dual push in models and applications
Tencent also formally released HunyuanWorld-1.0 at WAIC and made the model open source.
While the development path for large language models is increasingly defined—moving toward reinforcement learning—multimodal AI remains early in its evolution. Key hurdles include data sourcing, architectural design, and infrastructure challenges.
Multimodal capabilities have become a new competitive frontier in 2025. World models, which began gaining momentum in December 2024, are emerging as a crucial area of focus.
HunyuanWorld-1.0 integrates panoramic visual generation with 3D scene reconstruction. It can process both text and image inputs and generate stylized, explorable 3D environments. What once took professional teams weeks to build can now be produced in minutes using a single line of text or an image prompt.
Still, building robust world models is not straightforward. “Data remains a major bottleneck,” said Guo Chunchao, head of 3D development for Hunyuan, in an interview with 36Kr and other media outlets. “Most 3D assets are still handcrafted, and we only have tens of millions of samples to work with. That’s far below the volume of available image data.”
Looking ahead, Guo outlined two development goals. The first is improving the quality of 3D asset generation to reach commercial standards. Current results are promising but fall short of expectations in industries like gaming, autonomous driving, extended reality, animation, and film. Higher quality and broader generalizability would reduce costs and shorten production cycles.
The second goal is enhancing scene generation and interaction models to create more realistic simulations grounded in physical principles. Guo said this will be a focus throughout 2025, with the aim of achieving broader maturity by 2026.
The launch of DeepSeek-R1 earlier this year demonstrated how early entrants into new AI categories can gain influence and market share. Since then, more companies have open-sourced their models. Tencent followed with HunyuanWorld-1.0 and plans to release smaller-scale hybrid inference models by the end of the month, with parameter sizes ranging from 500 million to 7 billion. These lighter models are easier to deploy and fine-tune.
Tencent’s background in gaming and social platforms has helped position it as a strong contender in China’s multimodal AI race. Its open-source models are approaching commercial-grade performance, offering developers tools they can adapt for real-world use.
According to Tencent, its image and video derivative models now number 1,400 and 1,600, respectively. The Hunyuan 3D model series has surpassed 2.3 million downloads, making it one of the most widely used open-source 3D sets globally.
More releases are on the way. These include an edge-ready hybrid inference language model, a multimodal understanding model, and a visual generation model for gaming.
Among the upcoming offerings is Hunyuan-Large-Vision, a multimodal understanding model that currently ranks first in China on LMArena’s Vision Arena leaderboard. It is slated to be open-sourced soon. Another notable release is Hunyuan-GameCraft, a framework for interactive video generation in gaming environments.
Focus on deployment
Tencent has long taken a pragmatic approach to large model development. This year’s WAIC presentation continued that trend, emphasizing productivity as the central outcome of useful AI.
The company has embedded agentic features into a range of Tencent services, both consumer- and business-facing. These include tools for daily life, work, education, and entertainment.
In education, the QQ Browser’s QBot supports search, browsing, learning, and writing. The Ima workstation platform helps users manage routine tasks, build personal knowledge bases, and collaborate in real time.
Tencent has also developed a travel planning agent that can generate personalized itineraries with a single click and allow for adjustments on the fly. Through an integrated mini program, users can also make bookings, streamlining the entire process.
In music, QQ Music now offers AI-powered tools for songwriting and vocal synthesis, enabling users to create and perform songs using high-quality synthesized vocals.
Tencent is not just deploying AI agents. It is also building the infrastructure for others to do so. Its cloud-based agent development tools and Yuanqi platform lower the barriers to entry for both enterprise developers and individual creators.
Previously, industry-specific large models were often seen as expensive and difficult to implement. But as model capabilities continue to improve, agents are emerging as a more flexible way to bring AI into practical use.
Wu said that industry models and agents are not mutually exclusive. Instead, they are complementary. Companies can co-develop vertical models with Tencent Cloud to embed specialized knowledge into reusable systems. Agents can then be deployed at the application layer, using protocols like the Model Context Protocol (MCP) to solve specific problems.
“Agents amplify the value of large models,” Wu said. “They are a key approach to solving deployment challenges in the real world.”
KrASIA Connection features translated and adapted content that was originally published by 36Kr. This article was written by Deng Yongyi for 36Kr.