Author | Huang Yu

Editor | Zhou Zhiyu

Fierce battles are in full swing on the AI big model battlefield. Major domestic and foreign model makers, including OpenAI, Google, and ByteDance, have launched the latest AI products, and ByteDance even offered a price tag as soon as it debuted, hoping to get a share in this scuffle.

Tencent is not absent from this battle that affects the future. At the “Tencent Cloud Generative AI Industry Application Summit” on May 17, Tencent systematically showcased its big models and products to the outside world.

Wu Yunsheng, vice president of Tencent Cloud and head of Tencent Cloud Intelligence, did not directly respond in an interview after the meeting. He only stated that the underlying capabilities and construction tool chain of Tencent's big model are in the process of rapid iteration, so Tencent will focus a lot of energy on developing products and technical capabilities.

Generative AI has surged in the past year, and Tencent hopes to win the market through competitive big models and products. At this summit, Tencent focused on disclosing a series of developments in large-scale model research and development and industrial application products. An independent C-terminal application that had previously received much attention was also put on the table for the first time.

Jiang Jie, vice president of Tencent Group, revealed that Tencent Yuanbao, a C-side assistant app based on the mixed-yuan model, will be released on May 30. In the past, Hybrid Assistant only had PCs and applets. Yuanbao is the first AIGC application built by Tencent for the C-side.

However, as far as Tencent is concerned, C-side applications are still in the exploration stage. Currently, more energy is still focused on improving the capabilities of the underlying large model and industrial applications.

Since its debut in September 2023, Tencent's mixed-element model has expanded to a trillion dollar scale, and is the first in China to adopt a hybrid expert model (MoE) structure, which has upgraded the overall performance of the model while also saving computing power costs.

According to reports, some of the Chinese proficiency of the mixed yuan model is on par with GPT-4, and their ability to answer “current and new” questions has also improved in mathematics, reasoning, etc.

In terms of multi-modal capabilities, Tencent's mixed-element large model has also been improved. For example, in the field of raw images, Tencent's hybrid Wenshengtu infrastructure has been fully upgraded to the same DiT architecture as Sora; in the field of live video, Tencent mixed-element supports various video generation capabilities such as Wensheng video, TuSen video, graphic video, etc., and the video generation length can reach 16 seconds.

According to our understanding, these model capabilities will be opened to the public through the Tencent Cloud API to meet the model needs of enterprise customers and developers in different scenarios, so that large model capabilities can actually be implemented in industrial applications.

On Tencent Cloud, the mixed-element model provides model services of various sizes, such as hunyuan-pro with trillion parameters, hunyuan-standard with 100 billion parameters, and hunyuan-lite with 10 billion parameters. Among them, the hunyuan-standard model has an ultra-long contextual understanding ability of 256k, and the number of characters processed at a time exceeds 380,000.

Improving the capabilities of big models is certainly important, but unlike many big model manufacturers, Tencent has always used “industrial practicality” as the core strategy for developing big models. When the big model was released last year, Tencent emphasized that “building a big model is only the starting point; the goal is to implement technology into the industrial scene and create value.”

In order to lower the usage threshold for industrial customers, the distance from basic models to industrial applications is shortened. At this summit, Tencent also launched a new native toolchain for big models, including the big model knowledge engine, image creation engine, and video creation engine, to help enterprises improve the quality and efficiency of knowledge services, image and video creation.

Wu Yunsheng said that the purpose of developing the three major AI model engine tools is to simplify the processes of data engineering, model refinement, and application development, and help enterprises use large models more efficiently and conveniently.

Under the core strategy of “industrial utility”, a major development model, Tencent's path is clear, which is to build AI closest to the industry by building high-performance models, efficient tool platforms, highly agile scenario applications, highly usable computing power infrastructure, and a strong and secure model environment.

Before developing large-scale model capabilities externally, Tencent has chosen to empower its own business first. According to the latest data disclosed by Tencent, the mixed-element model has now been tested in more than 600 Tencent's internal businesses and scenarios.

In addition, Tencent Hybrid will also launch an intelligent ecosystem and launch “Tencent Elements”, a one-stop AI agent creation and distribution platform, which is equivalent to the Tencent version of “GPTS”. In the future, users will not only be able to create exclusive AI agents on the platform, use Tencent's official plug-ins and knowledge base, but also distribute these agents to channels such as QQ, WeChat customer service, and Tencent Cloud with one click.

If the created smart body can have a good chemical reaction in Tencent's social ecosystem, it may allow Tencent to harvest “new shoots” again.

The big AI model brings opportunities only once every few hundred years to all walks of life. On this “world's biggest” circuit, Tencent must wait hard to seize every opportunity to reinvent itself.