According to the Zhitong Finance APP, Jianghai Securities released a research report stating that on April 17, 2025, Volcano Engine will launch the Doubao 1.5 Deep Thinking Model, upgrade the Doubao Text-to-Image Model 3.0, and the Doubao Visual Understanding Model; at the same time, it will release the OS Agent solution and the GUI Agent large model—Doubao 1.5 UI-TARS model for Agent services; and for large-scale inference, it will release the AI Cloud Native Serving Kit inference suite. The firm continues to be optimistic about AI application investment opportunities and highlights key attention on HAND (300170.SZ), Dark Horse (300688.SZ), and INTSIG (688615.SH).
<h2>The main points of Jianghai Securities are as follows:</h2>
The daily tokens usage of the Doubao large model continues to rise sharply, benefiting the data element and computing power sectors
As of the end of March 2025, the daily tokens usage of the Doubao large model has exceeded 12.7 trillion, three times that of December 2024 and 106 times that of when it was first released a year ago. IDC reports show that in 2024, the usage of large models in China's public cloud surged, with Volcano Engine holding the top market share in China at 46.4%. The firm believes that the continuous rise in tokens usage of the Doubao large model is beneficial for the data element and computing power sectors.
The Doubao 1.5 Deep Thinking Model is newly released, adopting MoE architecture and a dual-track reward mechanism
The Doubao 1.5 Deep Thinking Model is newly released, performing excellently in reasoning tasks in professional fields such as mathematics, coding, and science, reaching or approaching the global top tier level; in non-reasoning tasks like creative writing, the model also demonstrates excellent generalization ability, capable of handling a wider range of complex scenarios. To enhance general capabilities, the model team optimized data processing strategies, integrating verifiable data with creative data to meet various task requirements. Large-scale reinforcement learning is a key technology for training reasoning models, and by adopting an innovative dual-track reward mechanism, it effectively optimizes the algorithm while balancing &#34;clear right and wrong&#34; and &#34;subjective opinions&#34; tasks. The model uses MoE architecture, with a total of 200 billion parameters and only 20 billion active parameters, providing significant advantages in training and inference costs. Based on efficient algorithms, the model offers extremely high concurrent capacity for the industry and achieves a very low latency of 20 milliseconds. When solving specific problems, the large model must be able to query internet information and conduct multi-round searches and thinking. Unlike other reasoning models that &#34;search first and think later,&#34; the Doubao APP is trained based on the Doubao 1.5 Deep Thinking Model to &#34;think while searching&#34;; in addition, this model also possesses visual understanding capabilities, allowing it to think based on what it sees. The firm believes that the innovation of the Doubao 1.5 Deep Thinking Model lies in its use of MoE architecture (total parameters of 200B, active parameters of only 20B) and a dual-track reward mechanism.
The Doubao Text-to-Image Model 3.0 is newly upgraded, with better text layout, image, and picture production effects
The Doubao Text-to-Image Model 3.0 is newly upgraded, capable of achieving better text layout performance, photo-level image generation effects, and 2K high-definition image generation methods; it can be widely applied in marketing, e-commerce, and design scenarios such as film and television, posters, painting, and doll design
In the latest authoritative ranking in the field of text-to-image generation, ArtificialAnalysis Arena, the Doubao Text-to-Image Model 3.0 has surpassed many mainstream models in the industry, ranking in the global first tier. The firm believes that the new upgrade of the Doubao Text-to-Image Model 3.0 is expected to be implemented in more application scenarios.
Doubao Visual Understanding Model upgraded, video positioning more accurate, video understanding more intelligent
The Doubao Visual Understanding Model has been newly upgraded, featuring stronger visual positioning capabilities, supporting bounding box positioning and point positioning for multiple targets, small targets, and general targets, as well as supporting positioning counting, describing positioning content, and 3D positioning. It can be applied in scenarios such as inspection of offline stores, GUI agent, robot training, and autonomous driving training. At the same time, the new version has also significantly improved video understanding capabilities, such as memory, summarization understanding, speed perception, and long video understanding. The Doubao Visual Understanding Model, combined with vector search, can directly perform semantic searches on videos, widely applicable in commercial scenarios such as security and home care. The firm believes that the new upgrade of the Doubao Visual Understanding Model is expected to continuously empower industries such as robotics, smart vehicles, and security in the future.
Facing Agent services, Volcano Engine releases OS Agent solution and GUI Agent large model - Doubao 1.5·UI-TARS model; for large-scale inference, Volcano Engine releases AI cloud-native Serving Kit inference suite
Volcano Engine believes that in the future, AI Agents will develop in parallel in two directions: application Agents and OS Agents. Application Agents have stronger specialization, such as customer service Agents, data Agents, code Agents, etc., capable of focusing on completing tasks in specific fields; while OS Agents possess cross-scenario universality and flexibility, able to directly operate browsers, computers, mobile phones, or other Agents to complete complex tasks. Based on this, Volcano Engine officially released the OS Agent solution. This solution encapsulates the capabilities of the Doubao large model through the Volcano Engine veFaaS platform, enabling enterprises and developers to easily build lightweight Codeuse and Browseruse. For complex OS Agents, Volcano Engine also officially released the GUI Agent large model - Doubao 1.5·UI-TARS model. This model integrates screen visual understanding, logical reasoning, interface element positioning, and operations into a single model, breaking through the limitations of traditional automation tools that rely on preset rules.
In addition, Volcano Engine launched the Serving Kit inference suite, assisting enterprises in achieving rapid deployment of models, inference optimization, and operational observability. The Serving Kit inference suite can complete the download and preheating of 671B Deep Seek R1 in 2 minutes and load the inference engine in 13 seconds. The firm believes that both application Agents and OS Agents will usher in rapid development in the future.
Risk Warning: Risks of changes in industrial policies, risks of AI application development not meeting expectations, risks of target company performance not meeting expectations

300170

300688

688615

Jianghai Securities released a research report stating that Volcano Engine has launched multiple Doubao models, including the Doubao 1.5 Deep Thinking Model and the OS Agent solution, and it is expected that AI applications will develop rapidly. The firm is optimistic about AI investment opportunities, focusing on HAND, Dark Horse, and INTSIG. The daily token call volume of the Doubao large model has reached 12.7 trillion, indicating strong growth, with Volcano Engine holding a 46.4% market share in China

Zhitong

Jianghai Securities: Volcano Engine releases multiple Doubao models, continues to be optimistic about AI application investment opportunities