ByteDance Released DouBao Large Model
On May 15, ByteDance officially released its DouBao large model. Currently, the DouBao large model processes 120 billion tokens of text daily and generates 30 million images.
At the press conference, Tan Dai, the president of Bytedance’s cloud service Volcano Engine, disclosed the commercial price of the DouBao large model — the main model of DouBao is priced at 0.0008 yuan per thousand tokens in the enterprise market, meaning it can process more than 1500 Chinese characters at a price of 0.8 cents, which is 99.3% cheaper than the industry average price. The price of models of the same specification on the market is generally 0.12 yuan per thousand tokens, which is 150 times the price of the DouBao model.
However, Volcano Engine did not disclose the specific parameter scale of the DouBao large model, because “the parameter scale is no longer the only standard to measure the ability of large models”. In an interview, Tan Dai said, “The industry is no longer competing for parameter scale this year, because everyone has realised [the fact that different sized models have different performances, and the prices are naturally different”.
In terms of specific cost optimization, Tan Dai revealed that on the one hand, the cost is reduced through technical means, such as optimizing and adjusting the model structure. Secondly, the past single-machine inference is upgraded to distributed inference, combined with the mixed scheduling of cloud computing power, to reduce costs in a comprehensive way.
As for why to make this market move of price reduction, Tan Dai stated that a big change this year is that the ability of large models in the industry has greatly improved, and application has become very important, that is, it is necessary to make the ecosystem prosperous. Tan Dai said that many clients he is currently in contact with are trying out large models, but the risk of innovation is very high, especially in the AI field, so it is necessary to reduce costs and encourage wider use. From this perspective, whether it is a large enterprise or an individual, everyone needs lower cost, higher quality large models.
At present, the application development of large models is still in the early stage. According to QuestMobile data, as of March this year, the number of users in the AIGC industry based on large models was 73.8 million, although it had grown 8 times year-on-year, it only accounted for 6% of mobile internet users. Tan Dai believes that reducing costs is a key factor in pushing large models quickly into the “value creation stage”.
Nowadays, the trend of reducing the price of large models in the industry is becoming more and more popular. Yesterday, OpenAI announced a price reduction while updating its large model technology. OpenAI’s two new embedded models: text-embedding-3-small and text-embedding-3-large, the former’s price is only one-fifth of the previous generation, the price per 1000 tokens has been reduced from $0.0001 to $0.00002. The latter can create up to 3072 dimensions of embedding, with its main improvement in performance.
So far, ByteDance has built the AI chat assistant “DouBao”, the AI application development platform “Button”, the interactive entertainment application “Cat Box”, as well as “Star Painting” and other AI creation tools based on the DouBao large model, and has integrated the large model into Douyin, Lark, as well as more than 50 other businesses to improve efficiency and optimize product experience.
Zhu Jun, Vice President of ByteDance’s Product and Strategy, believes that while technology is evolving rapidly, users’ core needs have not changed, including efficient information acquisition, work efficiency, self-expression, social entertainment, and so on. ByteDance hopes to find a suitable application form for large models. Currently, DouBao is a large model application that ByteDance is heavily investing in. According to Zhu Jun, there are already over 8 million intelligent entities created on DouBao, with a monthly active user base of 26 million.
Sign in with google
