ByteDance believes that the ToB and ToC markets for large models can evolve in tandem, fostering synergy.

At yesterday's Volcano Engine Force Conference, amidst a series of model releases and updates, ByteDance's AI strategy, particularly the synergy and tactics between ToB and ToC, garnered significant industry attention.

Currently, ByteDance boasts several C-end applications such as Douyin, Jinri Toutiao, and Maoxiang, with Douyin already enjoying substantial market traction. On the B-end, Volcano Engine offers external services.

For application development, ByteDance presents HiAgent for the enterprise market, aiding companies in constructing AI platforms that address both development and data integration challenges. Additionally, it introduces Kouzi, a no-code intelligent agent construction platform catering to a broader audience. Few other companies offer such distinct services.

As a late entrant in the ToB large model space, ByteDance is carving a new path into the ToB market by leveraging its C-end popularity to drive B-end engagement, thereby attracting attention and opportunities in the B-end sector.

01

Major Model Launches and Continued Price Reductions

The Douyin Large Model family witnessed significant updates yesterday with the release of two new models, upgrades to three existing models, and announcements for two upcoming launches.

The Douyin Visual Understanding Model represents a crucial step towards multimodal applications for the Douyin Large Model. Users can input both text and image-related queries, and the model can comprehensively understand and provide accurate responses.

From the demonstration video, it is evident that the Douyin Visual Understanding Model possesses various capabilities, including content recognition, understanding, reasoning, visual description, and creativity. Tan Dai emphasized that it would significantly expand the capabilities of large models, lower the barrier for interaction with large models, and unlock richer application scenarios.

Currently, the Douyin Visual Understanding Model is accessible to both ToC and ToB users through the Douyin App and PC products, as well as to enterprise customers via Volcano Engine.

Another newly unveiled model is Beaver3D, a 3D generation model supporting text-to-3D, image-to-3D, and multimodal 3D asset editing. This year, the embodied AI race has intensified, with giants like NVIDIA and prominent AI scientists in the industry pushing products and papers focused on embodied AI training. Domestic players are also accelerating embodied AI training from multiple perspectives, including data and algorithms.

According to a product developer at the exhibition booth, Volcano Engine has partnered with NVIDIA's Isaac robot development platform, becoming its sole cloud partner in China. The combined digital twin platform, veOmniverse, integrates with the text-to-3D model, enabling users to generate 3D scenes and models in real-time by simply inputting text, thereby meeting the diverse needs of simulation training.

Apart from the new models, ByteDance also enhanced the capabilities of three existing models: the large language model Douyin Pro, the music model, and the text-to-image model. It also announced the upcoming launches of a video generation model and an end-to-end speech model.

In addition to new model releases and upgrades, Volcano Engine continued its price reduction strategy.

The pricing model for the newly released Visual Understanding Model is still token-based, costing 0.003 yuan per thousand tokens. ByteDance claims that this price is 85% lower than the industry average. Meanwhile, the general model Douyin Pro also saw a substantial price reduction, costing one-eighth of GPT-4's usage price.

The reduction in model invocation costs has evoked mixed reactions among different groups.

For developers and enterprise users, the decrease in inference costs acts as a catalyst for the popularization of large model applications. An entrepreneur in the large model application space previously told DSZQ that the cost of a single invocation of ChatGPT was still several cents at the beginning of last year. For outbound call scenarios requiring model invocation, the business model was financially unviable. However, with the sharp drop in inference costs, an outbound call now costs only 0.2 cents in November this year, leading to an immediate acceleration in commercial applications.

In his keynote speech yesterday, Tan Dai also stated that reducing invocation costs can encourage enterprises to boldly innovate with large model applications.

From the perspective of large model vendors, previous price reductions have triggered a chain reaction within the industry. Well-known figures in the industry have publicly stated that increased competition will make it more challenging for some companies to survive.

Some large model application service providers have observed that ByteDance's pricing strategy serves as a lever to open up the ToB market, allowing Volcano Engine's intelligent agent construction platform HiAgent to be more rapidly promoted to enterprise clients.

02

Can ByteDance Succeed in ToB Business?

At the conference, ByteDance revealed application data for the Douyin Large Model. As of mid-December, the average daily token usage of the Douyin General Model has exceeded 4 trillion, a 33-fold increase from its initial release seven months ago.

Tan Dai believes that large models are thriving in various scenarios. For instance, in information processing scenarios, Douyin's invocation volume has increased by 39 times in the last three months; in customer service and sales scenarios, it has increased by 16 times; in hardware terminal scenarios, it has grown by 13 times, and AI tools have grown by 9 times. "ToB and ToC are progressing side by side," Tan Dai said in an interview.

"In the past, technology for ToB and ToC was segregated because the value of Douyin for consumers was completely different from the value of Volcano Engine for businesses. However, beneath the C-end and B-end of large models lies the same foundation," said Tan Dai. All capabilities are internalized within the model itself, fostering synergy between C-end and B-end.

A few days ago, information circulated widely in the industry that Douyin's user base was second only to ChatGPT. Some industry insiders believe that ByteDance, as a late entrant in the large model race, is making a strong impact, a trend that may prompt domestic competitors to adjust their strategies and potentially create opportunities for ByteDance in the ToB sector.

Tan Dai also mentioned that Douyin has garnered significant attention and new collaboration opportunities for Volcano Engine.

However, some industry insiders believe that ByteDance's strengths have always been in the C-end market, and it still needs to accumulate experience in the ToB sector of large models.

A large model application service provider previously told DSZQ that ByteDance had just started its large model privatization business and was still developing many strategies. They provided an example where HiAgent, similar to the privatized version of Kouzi, could not be privatized initially and focused on middleware capabilities, limiting the scope of cooperation from the perspective of application service providers. When enterprises initially adopted HiAgent but later struggled to utilize it and sought collaboration with application service providers, the service providers' own capabilities might not have been fully utilized. "Selling their own products may be ByteDance's current priority, and they have not yet fully considered the issue of systematization," said this service provider.

Yesterday, Tan Dai told DSZQ that a privatizable version of the Douyin model would be available on HiAgent. However, he believed that models evolve rapidly and that better models are hosted on the cloud, making it easier to conduct POCs and deployments based on the cloud.

Enterprises often prefer private deployments due to security and compliance considerations. Tan Dai noted that current technological advancements in Volcano Engine can effectively address many issues. For instance, the security intelligence solution achieves end-to-end encryption at the hardware level, similar to Apple's iPhones. Enterprises can leverage cloud-based models for technical convenience while meeting security and compliance requirements.

Volcano Engine also released a series of product updates to enhance its ToB service capabilities. Yesterday, Volcano Engine was updated with the introduction of a large model memory solution and an AI search and recommendation engine. HiAgent was also upgraded to version 1.5, featuring over 100 industry application templates derived from real-world enterprise scenarios, facilitating out-of-the-box and agile deployment for enterprises.

03

How ByteDance Synergizes Across Different Customer Segments

Yesterday's conference also provided insights into ByteDance's AI synergy between the ToC and ToB markets.

At the HiAgent booth, a developer inquired about the usage of the HiAgent platform, mentioning that they started using Kouzi's capabilities before moving to the professional version.

In the exhibition of application cases in the financial industry, it is evident that some enterprises began collaborating with ByteDance due to the Douyin App. For example, China Merchants Bank constructed an intelligent agent for mobile life offers on Douyin using Kouzi. This intelligent agent allows users to obtain restaurant recommendations with China Merchants Bank dining coupons through natural language dialogue, based on the LBS location provided by their mobile phones. Clicking on the coupons provided by the intelligent agent redirects users to the purchase interface for coupons on the China Merchants Bank mobile life app, completing the purchase and payment process.

"Douyin is now a significant traffic entry point due to its high monthly active user count. Therefore, enterprises place some scenarios on Douyin, such as dining and movie tickets. When users are ready to make a purchase, they can seamlessly complete the transaction through the redirect," emphasized the product technician at the booth regarding the traffic value of Douyin in such collaborations.

Moreover, ToB capabilities can also influence the quality of ToC services. At the medical exhibition booth, a solutions expert told DSZQ that they currently do not have an industry-specific large model but rather integrate medical knowledge and capabilities into the general Douyin model.

This medical knowledge corpus originates from Xiaohe Health (now known as Douyin Health), which was acquired by ByteDance previously. The medical knowledge incorporated into the base model meets the standards of the practicing physician examination. Additionally, the base model has been pre-processed to handle issues related to safety and advertising laws. It can recognize information on medical test reports, allowing enterprises to invoke it directly to provide ToC services.

At yesterday's conference, ByteDance also showcased industry solutions and implementation cases in key ToB sectors such as healthcare, finance, consumer goods, and education.

For example, in consumer scenarios, the AI search and recommendation capabilities of Volcano Engine enable users to discover products and content that better suit their needs. "When a user searches for a coat, they are not interested in seeing all coats but rather the ones they prefer. This scenario involves not only search but also recommendation capabilities. Not all data is suitable for recommendation and requires tagging. Large models possess strong comprehension abilities, enabling them to understand not only input data and tags but also descriptions and other modal content," explained an engineer at the exhibition booth.

Tan Dai mentioned that to better utilize large models, they established an algorithm and service team of over a dozen people early on to collaborate with clients on different scenarios and needs. He also emphasized the importance of building an ecosystem of partners.

DSZQ observed that Volcano Engine dedicated a special area for recruiting ecological partners at its conference booth, announcing the Douyin Enterprise Services Ecological Cooperation Plan. It is recruiting various partners, including SaaS service providers, AI-native innovative enterprises, AI integration service providers, data service providers, consulting firms, and technology open communities, focusing on technology sharing, product co-creation, market supply, and customer expansion.