Ping An Securities: Doubao large model family upgrade, productivity improvement and edge application potential highlighted

According to the Zhitong Finance APP, Ping An Securities released a research report stating that as AI large model applications penetrate and continuously unlock more new scenarios, AI is gradually moving towards the B-end and edge, which will also continuously demand updates and iterations for hardware terminals and related chips. Under the trend of intelligence, the upgrade of the large model family has improved performance and lowered the threshold, and the ByteDance Doubao large model is expected to accelerate the landing of AI on the edge. On the B-end, the Doubao large model helps improve enterprise productivity; on the C-end, the application potential of the Doubao large model on the edge is highlighted, and it is expected that the related industrial chain will usher in new opportunities.

Event: On December 18th, the ByteDance Volcano Engine Force Power Conference (Winter Meeting) opened, where the Volcano Engine brought a comprehensive upgrade of the Doubao large model family, releasing not only large models such as visual understanding but also upgrading several previously released large models.

The main views of Ping An Securities are as follows:

The conference announced an upgrade to the large model family, improving performance and lowering the threshold to promote the landing of large model applications.

Applications are the main way to monetize or realize the value of large models, and the market is also looking forward to good returns after large model manufacturers make large-scale capital expenditures. As one of the leading large models in China, the development and landing of the Doubao large model have also attracted significant market attention. At this Volcano Engine Force Power Conference, the company stated that since the release of the Doubao large model, user usage has rapidly increased, with the average daily tokens usage exceeding 4 trillion in December, an increase of more than 33 times compared to the release period in May. The conference officially announced the release of the visual understanding large model and 3D generation model, and also upgraded the general model pro, text-to-image model, music model, and others.

Specifically, the newly released visual understanding large model has stronger content recognition and visual description capabilities, and is expected to be applied in scenarios such as education, e-commerce marketing, and tourism. The conference announced a pricing of 0.003 yuan per thousand tokens, entering the "cent" era, and the low usage cost will significantly lower the usage threshold, which is beneficial for the promotion of the Doubao large model.

The company also made significant upgrades to the general model pro. As a large language model in the Doubao large model family, the capabilities of the newly released product have been greatly enhanced. Company data shows that the comprehensive capability of the model has improved by 32% compared to its first release in May, with professional capabilities improving by 54% and mathematical capabilities improving by 43%. Multiple indicators have aligned with GPT-4O, and in some complex scenarios, it has even surpassed it, but the usage price of the Doubao basic model pro is only 1/8 of GPT-4O. In addition, the upgraded text-to-image model has realized practical functions such as one-click photo editing and one-click poster creation. The continuous optimization of the Doubao large model capabilities is empowering the improvement of enterprise productivity.

On the B-end, the Doubao large model helps improve enterprise productivity.

Due to the continuous enhancement of capabilities, the Doubao large model has performed very well in key B-end areas, and the usage volume has also seen significant growth. According to the data released by the company at this conference, the usage volume in the information processing field has increased 39 times since its release, while the usage volume in customer service and sales has increased 16 times, and the usage in hardware assistant and AI tool scenarios has increased 13 times and 9 times, respectively At the conference, Volcano Engine also stated that the multimodal capabilities of the Doubao large model are rapidly improving. Its voice input function leverages its speed advantage to significantly enhance input efficiency and can import various voice signals, recognizing different language dialects and accents. Meanwhile, to promote the development of Doubao applications, Volcano Engine announced upgrades to three platform products: Volcano Ark, Button, and HiAgent, to help enterprises better build their own AI capability centers and efficiently develop AI applications.

On the C-end, the application potential of the Doubao large model is becoming prominent

In addition to serving as a productivity enhancement tool, the Doubao large model is actively developing on the edge side, and the prospects for AI applications at the edge are becoming clearer. At this release conference, Volcano Engine stated that the Doubao large model has already collaborated with 80% of mainstream automotive brands, with AI cockpit scenarios being one of the key cooperation directions. Automotive users can interact naturally, receive personalized services, and engage in multimodal interactions through the Doubao large model. On the mobile side, Volcano Engine is also actively integrating smart assistants, system controls, and creative tools into mobile scenarios.

Volcano Engine shared a case of the Honor smartphone application of the Doubao large model at the conference, highlighting its capabilities in areas such as magical photo editing. The company stated that Doubao has achieved significant results in terminal promotion, having developed over 50 application scenarios in the terminal field, covering approximately 300 million terminal devices. The token usage of the Doubao large model from smart terminals has increased 100 times in six months, showing very rapid progress.

Investment Recommendations:

(1) In terms of AI computing power, recommend Haiguang Information (688041.SH), Industrial Fulian (601138.SH), Inspur Information (000977.SZ), Unisplendour (000938.SZ), Zhongke Shuguang (603019.SH), and Digital China (000034.SZ), and suggest paying attention to Cambrian-U (688256.SH) and Jingjiawei (300474.SZ);

(2) In terms of SOC, recommend Hengxuan Technology (688608.SH), and pay attention to Espressif Technology (688018.SH), Xingchen Technology (301536.SZ), Zhongke Lanxun (688332.SH), Juchip Technology (688049.SH), and Allwinner Technology (300458.SZ);

(3) In terms of storage chips, recommend Zhaoyi Innovation (603986.SH);

(4) In terms of AI applications, strongly recommend Zhongke Chuangda (300496.SZ), recommend Kingsoft Office (688111.SH), iFlytek (002230.SZ), Foxit Software (688095.SH), Wankong Technology (300624.SZ), and Desay SV (002920.SZ), and suggest paying attention to Tonghuashun (300033.SZ) and Caixun Co., Ltd. (300634.SZ).

Risk Warning: 1) Application landing is not as expected; 2) Supply chain security risks; 3) Technology evolution and iteration are not as expected