Author | Three Norths
Edit | Yun Peng
Zhidong reported on May 15 that today, at the Volcano Engine Motive Power Conference, the ByteDance bean bag model officially opened its external services.
The Doubao large model family made its debut at the conference, and currently includes nine models: General Model Pro, General Model Lite, Role Playing Model, Speech Recognition Model, Speech Synthesis Model, Voice Replication Model, Wensheng Diagram Model, Function Call Model, and Vectorization Model.
In terms of pricing, Volcano Engine is pushing large models from "cent-based" to "cent-based" stages. The inference input price of the Doubao General Model Pro 128k version of the model is 0.005 yuan/1000 Tokens, which is said to be 95.8% lower than the industry price.
The inference input price of the Doubao General Model Pro 32k version is 0.0008 yuan/1000 Tokens, which is 99.3% lower than the industry price. In other words, one yuan can buy 1.25 million Tokens of the main model of bean bags, which is equivalent to three copies of "Romance of the Three Kingdoms".
▲ Tan Zhi, president of Volcano Engine, interprets the pricing of the main large model of bean bags
At the same time, Volcano Engine announced the launch of the Volcano Ark 2.0 platform, releasing three important plug-ins: networking plug-ins, content plug-ins, and knowledge base plug-ins. The Volcano Ark 2.0 platform can provide 10,000 calorie GPU resource pool support training, support the completion of kilocalorie expansion within 3 minutes, and improve the system carrying capacity, security and service capabilities.
In addition, Volcano Engine also announced the launch of Buckle Professional Edition, which provides an enterprise-level AI application development platform; Launched ChatBI, an AI assistant for data insights, ChatBI for intelligent creation, and Sales Copilot, an AI assistant. Cooperate with a number of industry partners to establish an intelligent terminal large model alliance and an automotive large model ecological alliance.
After the meeting, a few media such as Zhidong had a face-to-face dialogue with Tan Cheng, president of Volcano Engine.
What kind of thoughts and thoughts does the team have on the pricing of the bean bag model?
Tan said that the prices of models of different sizes and performance are different, and the pricing of the main model with the strongest ability of Byte this time is 99% lower than the industry price, which is very amazing.
There are two reasons behind it: first, the team can do it, and the team has a lot of optimization methods in technology, including the optimization and adjustment of the model structure to reduce the cost and achieve the effect, and greatly reduce the deployment cost through distributed inference and hybrid scheduling; The second is what the team needs to do, the application of large models has become more important this year, and the risk of large model innovation is still very high, so we need to reduce the cost of trial and error very low to be able to make large models widely used.
How is the performance of the bean bag model while the price is reduced?
Although the model parameters and benchmark performance were not specifically disclosed at the meeting, everyone will see a lot of third-party test results immediately after it is opened, and the team confidently accepts the evaluation after user use. The second is that ordinary users will have their own feelings after using the Doubao App, and its huge usage is also a good proof of the model's capabilities.
▲Volcano engine large model service full matrix diagram (Zhidong on-site shooting)
Tan revealed that the Doubao model currently processes an average of 120 billion Tokens text per day and generates 30 million images. The total number of downloads of the Doubao App has reached 100 million so far, and the number of monthly active users on both ends has reached 26 million.
1. Launched full-stack AI services, and opened external services for the bean bag model
Tan said that the development of large models is closely related to everyone and every enterprise, and the current implementation of large models by enterprises is facing key challenges such as model effect, inference cost and landing difficulty.
To this end, Volcano Engine announced the launch of a full-stack AI service to provide better models with lower costs and easier implementation, helping enterprises transform into AI.
At the same time, the bean bag model officially opened its external services. Tan revealed that after a year of iteration and market verification, the Doubao model currently processes an average of 120 billion Tokens text per day and generates 30 million images.
The Doubao General Model Pro has strong comprehensive capabilities such as comprehension, generation, logic, and memory, supports 128k long text fine-tuning, and supports rich scenarios such as Q&A, summarization, classification, and creation.
Lite is a more cost-effective general model, with the cost of 1,000 tokens reduced by 84% and the latency reduced by 50%, which supports scenarios such as small talk, weather and real-time information query, music and video playback, and navigation, and the effect is more than 50% ahead of traditional voice processing.
The Doubao role-playing model supports a high degree of character customization, can adapt to the needs of users to play or accompany based on contextual awareness, and has a strong ability to promote the plot and continuously guide the chat.
In the bean bag speech series model, the bean bag speech recognition model has high accuracy and personalization characteristics, and supports multilingual recognition. The bean bag synthesis model has the characteristics of natural, multi-emotional and multi-deduction. The bean bag sound reproduction model supports 5-second reproduction, cross-language migration, and has a high degree of similarity with the original sound.
In addition, the bean bag model also includes other subdivision models such as the Wensheng graph model, the function call model, and the vectorization model.
At the meeting, Zhu Jun, vice president of product and strategy of ByteDance, interpreted ByteDance's thinking on large-scale product design in combination with the case of bean bags.
Zhu Jun said that there are several basic principles in the design of bean bags, the first is to be anthropomorphic enough, for example, the name "bean bag" reflects anthropomorphic characteristics.
The second is proximity to the user. The functional positioning of Doubao includes the "voice know-it-all" that you can carry, the desktop copywriting assistant, etc., which are mainly embedded in the user's existing use environment to narrow the distance with the user.
The third is to follow personalization. In the world of agents, everyone has individual needs and emotions, so on the bean bag, each ordinary user can "pinch" an agent of their own. Doubao also supports the creation of higher-order agent definition capabilities, such as creating personalized tutors.
Up to now, the total number of downloads of the Doubao App has reached 100 million, the number of monthly active users on both terminals has reached 26 million, and the total number of agents created has reached 8 million.
2. The Volcano Ark 2.0 platform was released, supported by the Vanka GPU resource pool, and the three major plug-ins were upgraded
Today, Volcano Engine also announced the launch of the Volcano Ark 2.0 platform, releasing three important plug-ins: networking plug-ins, content plug-ins, and knowledge base plug-ins.
Among them, the networking plug-in supports real-time networking sources, provides the same search capabilities as Douyin and Toutiao, and supports multi-modal interaction and intent recognition retrieval.
The content plug-in supports massive content retrieval, and the Douyin content plug-in is exclusively on the shelves, providing Douyin rich video and graphic content, enriching the interaction process between large models and users, and supporting content strategy customization.
The platform knowledge base plug-in supports enterprises to call the internal knowledge base, and has a built-in search engine independently developed by Bytes, which supports millisecond-level and tens of billions of scale retrieval, and the search update is fast and the search relevance is high.
In addition to plug-in upgrades, the Volcano Ark 2.0 platform has improved its system carrying capacity, security and service capabilities.
Tan said that the platform can provide a 10,000-calorie GPU resource pool to support training, complete the expansion of 1,000 calories within 3 minutes, enhance the O&M experience and security compliance, and help the last mile of AI scenarios.
3. The professional version of the button was released, the AI assistant ChatBI was launched, and the intelligent creation cloud 2.0 was upgraded
Today, Volcano Engine launched the Buckle Professional Edition, which provides an enterprise-level AI application development platform.
It is reported that the button is a new generation of AI application development platform launched by Byte, which has the characteristics of low threshold, personalization, real-time and multi-modality, with massive AI resources, rich release channels, and support one-click custom API services.
At present, the professional version of the buckle has been integrated on the large model service platform "Volcano Ark" of the volcano engine. China Merchants Bank, Haidilao Hot Pot, Super Orangutan, Liepin and other enterprises have built intelligent bodies on the buttons. Fudan University, Zhejiang University and other famous universities have also set up AI "teaching assistants" for courses and experiments.
Today, Volcano Engine Intelligent Data Insight DataWind officially released the AI assistant ChatBI, which allows users to generate indicators through natural language interaction for chart making, data query, and in-depth data analysis.
For marketing scenarios, Volcano Engine has upgraded Intelligent Creation Cloud 2.0 to become a one-stop enterprise content marketing growth solution from content creation, matrix distribution, advertising to data insight. It not only provides multi-modal understanding and generation capabilities, helping to improve the efficiency of video creation by 25 times, but also launches marketing tools such as Douyin topics, POIs, and mini programs, increasing store page exposure by 600%.
For sales scenarios, Volcano Engine has released Sales Copilot, a sales AI assistant, which can efficiently follow up customer needs and assist sales in answering complex product questions at any time. In addition, it can also simulate different styles of customer rehearsals through role-playing, and provide excellent speech learning to improve the quality of sales communication.
In addition, Volcano Engine today announced the establishment of a smart terminal large model alliance with OPPO, vivo, Honor, Xiaomi, Samsung, and Asus; With more than 20 manufacturers such as Geely Automobile, Great Wall Motor, Jietu Automobile, Celis, and Zhiji Automobile, it announced the establishment of an automobile model ecological alliance.
Conclusion: A new price war in the large model industry has begun
At present, the price war of global large-scale model production has begun. At that time, OpenAI's latest flagship model, GPT-4o, has just announced that the API pricing will be cut in half; At this time, the domestic ByteDance bean bag model will blow up the price, and the main model is 99% lower than the industry price.
With the debut of the bean bag large model family, we see that the byte large model has the characteristics of strong application orientation. Guided by application requirements, ByteDance and Volcano Engine pay attention to the balance between multiple dimensions such as model effect, cost and ease of use. One of the major features of this bean bag model is that it is much lower than the pricing in the industry market, and we continue to pay attention to further feedback from users and developers after the experience.