01. Breaking down the floor price, what confidence does the byte have?
The plot of "breaking down the floor price", when the industry enters the white-hot competition, is inevitable after all, and it only depends on who makes the first move. Suddenly, Byte, which had been a little sluggish, chose to be a late striker and directly hit the calculation cost of the large model to an "unbelievable" cheaper.
On May 15, ByteDance's Volcano Engine announced at the "FORCE Power Conference" that the previously announced lark model was renamed and upgraded to "Doubao", and released 9 bean bag large models "Family Family Bucket" in one go.
In the face of the situation that giants at home and abroad have staged the "AI Technology Spring Festival Gala" one after another and fought in turns, Byte, which once chose to "wait and see", still came up with a price killer.
At the press conference, when showing the bean bag model family, Tan Cheng, president of Volcano Engine, did not even mention how many parameters and performance information of the bean bag, and showed the price as soon as it came up: 0.0008 yuan / thousand Tokens, indicating that this is a "floor price" lower than 99.3% of the industry.
(Photo source/Volcano Engine)
This is true, this is equivalent to the processing capacity of 2 million Chinese characters for the general model of bean bags that can be bought for 1 yuan.
Taking OpenAI's new model GPT-4o as an example, its main focus is that it is cheap and easy to use, with a price of $5 / million tokens (about 0.035 yuan/1000 tokens) and an output of 15 US dollars/million tokens (0.1 yuan/1000 tokens).
The current average price of domestic large-scale model manufacturers is about 0.12 yuan/1,000 Tokens - and the price of the bean bag model is much cheaper than the above-mentioned competition, which is equivalent to the calculation cost of 1500 Chinese characters is 0.8%. Compared with the same parameters, the price of bean packets is 0.2% of the GPT-4-32k fee; Baidu Wenxin Yiyan and Ali Tongyi Qianwen charge about 0.6%.
In addition, the renamed Doubao large model, in addition to supporting the to C version of the App, has also officially opened to the outside world to B and to the developer services, upgrading the previous lark large model to Pro and Lite.
Among them, the Pro version is the professional version of the LLM model developed by ByteDance, compared with the hottest Kimi and other long text players, the window size supports up to 128K long text. Lite is a more cost-effective lightweight version, which is more economical than the Pro version, with an 84% lower cost per 1,000 tokens and a 50% lower latency.
With the price magic weapon in his hand, Tan Zhi was full of confidence at the press conference. He emphasized that this price reduction is the main model price reduction, not the use of small model price reduction to confuse the public; And shouted directly to Ali and Baidu: "It's a pity that Ali and Baidu don't have a 128k fine-tuned model yet, but I'm looking forward to their updates." ”
02, byte sprint, no Buddha
Will this byte price cut set off a bloody "vicious competition" in the industry? At the press conference, Tan Bei's explanation for this is: the price of large models is reduced, not only low-cost lightweight versions can be provided, but also the main models and the most advanced models must be cheap enough.
From the perspective of the industry, it is likely to break the cost price war, which of course means that Byte's massive capital investment also marks Byte's determination to break out of the encirclement.
In 2023, when the market is the hottest, the large-scale model action of the volcano engine once showed "Buddha". In March and April 2023, Baidu Wenxin Yiyan and Ali Tongyi Qianwen were announced successively. In August 2023, Byte launched the AI chatbot "Doubao" for the first time; In September, the Byte Skylark model surfaced.
One of the reasons behind the "slow" action is that Byte's technical accumulation in AI large models is relatively weak. According to the analysis of industry insiders, Byte has a huge amount of short video traffic as a confidence, and how to apply the AI large model to adapt to short videos, there are both confidence and doubts.
According to media reports: At ByteDance's annual all-staff meeting held in January 2024, ByteDance CEO Liang Rubo once mentioned that the company did not start discussing GPT until 2023, and the industry's leading large-scale model startups were founded from 2018 to 2021.
In February 2023, Byte deployed a team codenamed "seed" on the large model, focusing on the model layer, led by Zhu Wenjia, who has served as the CEO of Toutiao and the head of product technology at TikTok.
According to "Phoenix Technology", at the end of 2023, Byte's AI layout will accelerate sharply: the Flow department has been officially established to focus on the application research and development of AI large models. In March 2024, Qi Junyuan, vice president of product at Feishu, was transferred to the Flow department, mainly responsible for the PC business of Doubao. In April 2024, Zhu Wenjia will be in charge of the overall byte AI business, reporting to Liang Rubo.
After the strategy accelerated, a period of sprint began within Bytes. A product person from Doubao revealed to "Phoenix Technology" that in Q1 2024, Byte's AI business (mainly referring to the Flow department) has been maintaining high-intensity operations, "basically all of them are single off, but the data performance of several products is not bad, especially Doubao, and the internal morale is booming."
After Byte upgraded its AI strategy, it superimposed strategies such as traffic support, and application DAUs such as Doubao also grew rapidly. At the end of 2023, according to media statistics, the monthly active users of Doubao are far inferior to Wenxin Yiyan and Tongyi Qianwen. Around March this year, according to the latest report from QuestMobile, in March 2024, the monthly active users of the "Doubao" App will be 23.282 million, ranking first among AIGC applications.
Stepping up customer acquisition has also been put on the agenda by bytes. A corporate user told "City Boundary" that Volcano Engine is in close contact with it recently: "What we do is a legal technology product, and we are connected to the API of Zhipu AI. Volcano has been coming to me frequently lately and wants me to switch to bean bags. ”
However, from the perspective of the market, there are still many uncertainties in the way low-price play can help bean bags win the minds of users.
A sales person of a large model company saw that from January to April 2024, more than 200 domestic large model winning projects were awarded, and the products of the large model are undergoing rapid testing by the market. He told "City Boundary": "The price of large model rolls is a good thing, but 'affordable' is not the same as 'good use', what customers really care about is the effect, performance, response speed, etc., and the low price regardless of cost is unsustainable." ”
Author|Dong Wenshu
Edited by Li Yuan