Following the recent launches of artificial intelligence products by OpenAI, Google, and ByteDance, Tencent has finally entered the scene. On May 17, Tencent unveiled several updates to its suite of AI products, including enhancements to its foundational models, industry-specific model upgrades, and the launch of multiple new products.
Among the highlights was the launch of a new product called Yuanqi.
Essentially, Yuanqi is Tencent’s version of GPTs, allowing enterprises and developers to create digital agents using official Tencent plugins and knowledge bases. Once developed, these agents can be conveniently utilized across QQ, WeChat, Tencent Cloud, and other compatible channels.
Conversation has, to date, been the most direct method for end users to communicate with large models. Since ChatGPT sparked the generative AI wave, the industry, particularly in China, has closely watched Tencent’s progress with large models and their integration into tech ecosystems like WeChat and QQ. Yuanqi’s launch offers some insights into these questions.
However, drawing from past experiences, OpenAI transitioned from third-party ecosystems via GPT plugins shortly after its release to launch the GPT Store in November 2023, offering millions of GPT options upon launch. Many of these were simple wrappers of ChatGPT—easy to develop but with less than ideal functionality. Similarly, Tencent will face significant challenges if it wants to build a robust ecosystem with Yuanqi.
Another important announcement is that Tencent will launch a new assistant app called Yuanbao by the end of May, marking a significant entry-level release. Previously, retail users for Tencent’s Hunyuan were largely limited to the Hunyuan Assistant mini program and the website app, without a unified mobile entry point. Yuanbao’s launch is expected to address this gap, although Tencent did not provide detailed information about the new product, only briefly introducing its features such as search, translation, document summarization, and spoken language practice.
Tencent’s announcement of the app signals its intent to compete in the consumer-facing AI product space. However, making an impression won’t be easy. Fellow tech giants Baidu and Alibaba have already introduced Ernie Bot and Tongyi Qianwen, while startups like Moonshot and Zhipu AI have also launched their own offerings, Kimi and Zhipu Qingyan, respectively. Tencent should therefore expect considerable turbulence as it attempts to make headway with Yuanbao.
Updates to Tencent’s large models
Since the release of its foundational Hunyuan large model last year, Tencent has maintained a steady pace of updates. The current Hunyuan model, which adopts a mixture of experts (MoE) structure, has improved its overall performance by 50% compared to the previous generation, with some Chinese language capabilities said to be on par with GPT-4.
Various capabilities have finally caught up, such as the “long text” feature that major companies and startups have been competing over. In this release, Tencent has also officially launched the 256K version of Hunyuan, capable of processing ultra-long texts exceeding 380,000 characters. In long text input scenarios, Hunyuan’s accuracy rate for the “needle in a haystack” test can reach up to 99.9%.
In application, if you input the contents of a book like “Romance of the Three Kingdoms,” which has hundreds of thousands of words, the model can identify key characters and events and provide precise information on details such as weather and character attire. In conversational scenarios, the model is better able to recall dialogue content and provide more accurate feedback by analyzing the context to assist in decision-making.
Currently, Hunyuan offers three model sizes: Pro, Standard, and Lite, available for both enterprises and individual developers.
On May 14, Tencent announced another significant release: the full upgrade and open-sourcing of the Hunyuan text-to-image model, including model weights, inference codes, and algorithms, allowing free commercial use by enterprises and individual developers. This is the first native Chinese open-source text-to-image model, adopting the same diffusion transformer (DiT) architecture as OpenAI’s Sora, serving as a fundamental model for text-to-image and text-to-video applications.
Start internally before expanding outward
Since Hunyuan’s official launch in September 2023, Tencent has pursued a more stable route in the AI field, starting with internal businesses and pushing to the industry only after capabilities mature. Internally, Tencent’s adoption of AI has been rapid. At the beginning of 2024, a senior Tencent representative told 36Kr that over 300 internal businesses had adopted Hunyuan. Today, this number is believed to surpass 600.
For example, Tencent’s AI code assistant has already been adopted by more than 50% of the development staff within the group, with a code generation rate exceeding 30% and a more than 20% improvement in R&D efficiency. Tencent’s internal ecosystem spans content, social, and gaming businesses. With the support of large models, many businesses have made significant progress.
For instance, a new feature in WeChat Read, based on the Hunyuan large model, enables users to ask AI about the themes of books without needing to read them entirely. Other Tencent products have also seen considerable growth with large model capabilities. The AI assistant feature in Tencent Meeting is a good example, offering opportunities to improve efficiency by performing tasks such as speech reminders, viewpoint summaries, and meeting minutes through simple commands. Over the past four months, daily usage of the feature has increased 20-fold, according to 36Kr.
Now, Tencent is gradually shifting externally, accelerating its pace of industrial implementation. A clear sign is that Tencent Cloud released three new engines at the platform-as-a-service (PaaS) level: the large model knowledge engine, the image creation engine, and the video creation engine. For instance, the knowledge engine was used by an insurance company called Yuan Xin Insurance Services to develop an efficient assistant for insurance agents, capable of automatically generating product knowledge Q&A and conversation scripts with clients, ostensibly improving efficiency by up to 50%.
These capabilities existed in the cloud computing era, but their combination with large models covers more scenarios. For example, based on Hunyuan’s video generation capabilities, users can input videos to immediately generate videos in specific styles. The generated video frames are smooth and natural, with strong temporal consistency. These capabilities can be output as APIs, enabling developers to use Tencent’s development platform to create more feature-rich applications.
Overall, major domestic companies are heavily investing in AI, advancing in both B2B and B2C directions, but there are already subtle distinctions in their specific routes. ByteDance, which is also making cautious progress, appears to be more B2C-oriented based on the products released. This is evident from using the name Doubao for both the large model and the app. ByteDance also continues to use the “app factory approach,” developing a large number of B2C apps based on Doubao.
Tencent, on the other hand, is taking a more industry-oriented route, having implemented its large models in over 20 industries, including finance, healthcare, education, automotive, and energy.
This article was written by Yong Yi in Chinese and was originally published by 36Kr.
