ByteDance started a price war for large models? The price is 99.3% lower than that of its peers, saying that there is no need to exchange losses for income

The Paper

2024-05-15 22:24Published on the official account of Shanghai Paper

"It's not sustainable to trade losses for income, so we never go down that path."

On May 15, ByteDance officially released the bean bag model at the Volcano Engine Power Conference. Volcano Engine is a cloud service platform under ByteDance, and Tan Zhi, president of Volcano Engine, talked about the pricing strategy of large models in an interview with The Paper and other media.

Tan Bei introduced that after a year of iteration and market verification, the bean bag model is becoming one of the largest and most abundant application scenarios in China, with an average daily processing of 120 billion Tokens text and generating 30 million pictures.

The price is the biggest highlight of this launch. According to reports, the pricing of the main model of bean bags in the enterprise market is only 0.0008 yuan/1000 Tokens, and 0.8 centimeters can process more than 1,500 Chinese characters, which is 99.3% cheaper than the industry.

It is worth noting that as early as August 17 last year, ByteDance had quietly announced that Doubao would open public beta, but it was delayed in showing up to the public half a year later. In this regard, Tan Cheng said frankly that in order to achieve a positive cycle for domestic enterprises to expand their models, they not only hope that the technology will mature, but also continue to optimize the basic performance and scenario implementation through ecological construction and more practices, provide more successful experiences, and in turn drive technological progress.

Large-scale models start a "price war"?

The bean bag model, formerly known as "Skylark", is one of the first large models in China to pass the algorithm filing. Regarding the pricing of bean bags, the outside world generally believes that this is ByteDance's first "price war" in the field of large models.

The pricing of the bean bag model is significantly lower than the industry price: taking the pro-32k version of the bean bag general model as an example, the input price of the model inference is 0.0008 yuan/1,000 tokens. The pricing of the model of the same specification on the market is generally 0.12 yuan/1,000 Tokens, which is 150 times the price of the bean bag model.

According to the price announced by Volcano Engine, one yuan can buy 1.25 million Tokens of the main model of bean bags, which is about 2 million Chinese characters, which is equivalent to three copies of "Romance of the Three Kingdoms".

Regarding pricing, Tan Cheng admitted, "There are two main reasons behind it: the first is that we can do it, and the second is that we need to do it." He introduced that the Volcano Engine has a lot of optimization methods in technology, which can reduce the cost of large models, and in addition, the underlying computing power can be better used in the form of distributed inference. In addition, in terms of hybrid scheduling, different loads can be used as hybrid scheduling, which also greatly reduces the cost.

Tan said that this year is the year of the promotion and application of large models, and it is necessary to establish an AI-related ecology, but the risk cost of innovation is very high, and only by reducing the cost of trial and error can a wider range of popularization be achieved.

"It's not sustainable to exchange losses for revenue, so ByteDance has never taken this path." Tan said that the main reason behind the price reduction is that it can be achieved technically, and there are many ways to continue to reduce costs in the future, "We are still confident in technology, which is the most critical point, and losses are unacceptable." ”

For the competition in the large model market, Tan said that the current large model market is still in the early stage of development, "how big the pond is, how big the fish are, and now the whole pond is still in the process of cultivation." "At present, the large model is still in a state of contention, and everyone needs to work together to achieve application landing, reduce costs, and benefit more people, so there is no fierce competition at present.

It is worth noting that OpenAI's latest release of GPT-4o, as a representative enterprise of China's large model, how does ByteDance view the technological gap between China and the United States? In this regard, Tan said that in terms of seeking truth from facts, OpenAI is still the first in the world, and the gap must be acknowledged. However, ByteDance's large model is still evolving, and it can now achieve wide application popularization, "Although there is a gap, the speed we are chasing is still very fast, and the native application, experience and user volume of Doubao are constantly improving." ”

As for the strong alliance between mobile phone manufacturers and large models that the outside world is concerned about, Tan Cheng also expressed his expectations: "This year is the first year of AI mobile phones, and theoretically through AI applications can achieve a better experience of mobile phone interaction, which requires cloud manufacturers, model manufacturers and mobile phone manufacturers to do more co-creation and in-depth cooperation." ”

ByteDance's AI play

As a major domestic technology company, ByteDance's AI roadmap has always been quite mysterious, and it is quite cautious about public releases to the outside world.

According to the data, the Doubao model provides a model family with multimodal capabilities, which currently mainly includes nine models, including general model Pro, General Model Lite, speech recognition model, speech synthesis model, and Wensheng diagram model.

Based on the Doubao model, ByteDance has created an AI dialogue assistant "Doubao", an AI application development platform "Button", an interactive entertainment application "Cat Box", as well as AI creation tools such as Star Painting and Instant Dream, and connected the large model to more than 50 businesses such as Douyin, Tomato Novels, Feishu, and Giant Engine to improve efficiency and optimize product experience.

In February this year, with the advent of Sora Wang Bomb, it was rumored that ByteDance was already developing a "Chinese version of Sora": an innovative video model called Boximator before Sora detonated the Wensheng video track.

At that time, a relevant person from ByteDance told The Paper that Boximator is a research project on technical methods to control the movement of objects in the field of video generation. At present, it cannot be implemented as a perfect product, and there is still a big gap between it and the leading foreign video generation models in terms of picture quality, fidelity, and video duration. ByteDance researchers said that the model is still in the research and development stage, and it is expected to release a test website in 2-3 months.

According to public information, as early as 2016, Byte established an artificial intelligence laboratory AI Lab, focusing on natural language processing, machine learning, data mining and other aspects of research.

At the end of 2023, Byte officially established the Flow department, focusing on the application research and development of AI large models. Jun Zhu, vice president of product and strategy at Byte, and Dingkun Hong, vice president of Byte Technology, serve as the product leader and technical leader of the department, respectively. In April, Byte promoted the head of the AI department, and now Zhu Wenjia is in charge of ByteDance's AI business as a whole, reporting to Liang Rubo, CEO of ByteDance.

Although the disclosure is relatively low-key, ByteDance's AI report card is quite bright: the data shows that in the Apple App Store and major Android app markets, the number of downloads of the Doubao APP ranks first among AIGC applications. More than 8 million agents have been created on Doubao, and the number of monthly active users has reached 26 million.

View original image 67K

•