According to a research report released by Open Source Securities, based on data from Volcano Engine, the daily average tokens usage of the Doubao large model exceeded 4 trillion in December, increasing more than 33 times compared to the time of its release in May. The invocation volume of the Doubao large model in scenarios such as information processing, customer service and sales, hardware assistants, and AI tools is also growing rapidly. The continuously increasing invocation volume and multi-scenario coverage make the Doubao large model more comprehensive and welcome a full upgrade, with the Doubao general model pro completing a new version iteration, enhancing its comprehensive task processing capability by 32% compared to May.

KYG Securities' main points are as follows:

ByteDance releases the Doubao visual understanding model, with multi-scenario applications expected to take off.

On December 18, Douyin's Volcano Engine, a subsidiary of ByteDance, released the Doubao visual understanding model at the 2024 FORCE original power conference. With the Doubao visual understanding model, users can input text and image-related questions simultaneously. The model can comprehensively understand and provide accurate answers, greatly simplifying the development process.

The Doubao visual understanding model mainly has three capabilities: (1) stronger content recognition capability, which can not only identify the categories and shapes of objects in images but also understand the relationships between objects, spatial layout, and the overall meaning of the scene; (2) stronger understanding and reasoning capability, which can identify text and graphic information and perform complex logical calculations; (3) more delicate visual description capability, which can provide more nuanced descriptions of the content presented in the images based on image information and can create in various styles.

Based on these abilities, the Doubao visual understanding model has extensive applications in scenarios such as Education, tourism, and e-commerce. For example, in the education scenario, it optimizes essays and popular science knowledge for students; in tourism, it helps tourists read foreign language menus and explains the background knowledge of the buildings in the photos; in e-commerce marketing scenarios, it helps merchants fully describe product details and efficiently publish promotional advertisements, etc. Moreover, the input price for Doubao visual understanding is 0.003 yuan per thousand tokens, which is 85% lower than the industry average price, making it favorable for businesses and developers to utilize the visual understanding model and create commercial value in broader scenarios.

The usage of the Doubao large model has significantly increased, with the model family expanding and multi-modal capabilities continuously improving.

According to Volcano Engine data, the daily average token usage of the Doubao large model exceeded 4 trillion in December, growing over 33 times since its release in May. The usage of the Doubao large model in scenarios like information processing, customer service and sales, hardware assistance, and AI tools is also rapidly increasing. The continuously rising usage and multi-scenario coverage make the Doubao large model more comprehensive, leading to a complete upgrade, wherein the Doubao general model pro completed a new version iteration, improving its overall task handling capability by 32% compared to May.

In addition to the visual understanding model, Volcano Engine also released Doubao music model 4.0, Doubao text-to-image model 2.1, and veOmniverse + Doubao 3D generation model. The Doubao video generation model will officially open for service in January 2025, and next spring, ByteDance will release Doubao video generation model version 1.5 with enhanced capabilities for generating longer videos. Additionally, Volcano Engine launched an all-domain AI search, integrating scene-based search recommendation services, enterprise private domain information consolidation services, and connected Q&A services, closely linking enterprise information, business, and user needs, accelerating intelligent transformation across multiple industries.

The Doubao large model is expected to drive rapid development in the large model industry, focusing on investment opportunities related to AI applications.

Training of the Doubao multi-modal model may drive demand for textual, image, and 3D material corpora. Key recommendations include Funshine Culture Group, with beneficiary symbols including Visual China Group (000681.SZ), Silkroad Visual Technology (300556.SZ), Tianyu Digital Technology (002354.SZ), COL Group Co.,Ltd. (300364.SZ), and IReader Technology (603533.SH). The Doubao music model may accelerate penetration of AI music users, with key recommendations for Hubei Century Network Technology Inc. (300494.SZ). The Doubao video generation model may accelerate the production of film and television content and improve efficiency in IP monetization, with key recommendations for Shanghai Film (601595.SH), and beneficiary symbols including Beijing Jetsen Technology (300182.SZ), Zhejiang Huace Film & TV (300133.SZ), and Beijing Enlight Media (300251.SZ).

The Doubao visual understanding model may accelerate the commercialization of multi-scenario AI applications: AI + e-commerce/marketing, with key recommendations for MOBVISTA (01860) and Beijing Zhidemai Technology, with beneficiary symbols including Inly Media Co., Ltd (603598.SH), Guangdong Insight Brand Marketing Group (300781.SZ), and Foshan Yowant Technology (002291.SZ). AI + companionship/toys, with key recommendations for Alpha Group (002292.SZ), benefiting symbols including Zhejiang Jinke Tom Culture Industry (300459.SZ). AI + education Publishing, benefiting symbols include Astro-century Education & Technology (300654.SZ), Beijing Shengtong Printing (002599.SZ), and Southern Publishing and Media (601900.SZ).

Risk warning: The iteration speed of the Doubao large model may not meet expectations; the commercialization process of Doubao AI applications may not meet expectations.