news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Source: Cover News

On July 25, the first stop of Volcano Engine 2024 "AI Innovation Tour" Chengdu Station officially set sail! If you don't sail, you'll be amazing. It is reported that as of July 2024, the average daily usage of Tokens of Doubao Big Model exceeded 500 billion. Since its release two months ago, the average daily usage of Tokens of each corporate customer has increased by 22 times.

This tour will release the Doubaotu model for the first time, as well as upgrade the Wenshengtu model, speech synthesis model and sound reproduction model. In the Doubaotu model, you can convert your photos into the most popular clay style and Monet style with one click. In the application of the sound reproduction model, you can actually hear "Taibai Jinxing" speaking a foreign language online. All of these make people have to sigh at the powerful functions of the Doubao model.

In terms of price, Doubao Big Model's "stronger model, easier to implement, and lower price" is not just talk. It should be convenient for enterprises to trial and error, and powerful. Doubao Big Model family has rich models and diverse application scenarios. It helps enterprises to easily build high-quality AI applications at low cost, driving business growth while bringing innovative business experience.

1. Large model capability test is "far ahead":

The so-called "three-strong competition" means standing out in terms of model, price and practicality. The quality of a large model lies in its usage, and only with large usage can a better model be polished. The Doubao large model has been continuously polished under the "hardening" of hundreds of billions of daily tokens, and has been widely recognized in terms of its capabilities and reasoning effects.

According to the latest evaluation list released by FlagEval, a large model evaluation platform under Zhiyuan Research Institute in June, the list shows that in the "objective evaluation" of closed-source large models, Doubao-Pro-4k won the first place among domestic large models.

Doubao Big Model is the first to help enterprises conduct low-cost trial and error with the strongest version, stronger model and lower price. For example, the Doubao Universal Model Pro, which is on the list this time, has a 32k version of the model inference input price of only 0.0008 yuan/thousand tokens. Simply translated, just like processing the 750,000-word text volume of "Romance of the Three Kingdoms", it only costs 1 yuan to process 3 books, and processing the 8 books of the Chinese version of "Harry Potter" (2.74 million words), it only costs 1.5 yuan and there is still some left. Even in the case of such fierce competition in the large model market, Doubao has achieved the ultimate in cost performance.

2. Domestic large-scale bean bags have the ability to make money:

According to Zhang Xin, vice president of Volcano Engine, more than 50 businesses within ByteDance are using Doubao Big Model, covering various scenarios such as collaborative office, data analysis, copywriting, auxiliary programming, content review, customer service, game NPC, role dialogue, education, etc. The new technology engine built based on Doubao Big Model is accelerating business innovation. In addition, Doubao Big Model's external customers have covered more than 30 industries such as mobile phones, automobiles, finance, consumption, and interactive entertainment. It has also established the Smart Terminal Big Model Alliance and the Automobile Big Model Ecological Alliance in cooperation with well-known terminal manufacturers such as OPPO, vivo, Honor, Xiaomi, Samsung, and Asus, and more than 20 automobile manufacturers such as Geely Automobile, Great Wall Motors, Jetour Automobile, Seres, and Zhiji Automobile.

For example, OPPO has explored and invested in multiple AI service scenarios such as emotional companionship, content creation, encyclopedia travel, and smart office. In addition, many well-known universities have also used related technologies to build "AI assistants", "library book search" and "Cyberspace Administration service assistants" for courses and experiments.

As typical corporate customers of Doubao Big Model in the southwest region, the blue-collar recruitment platform Yupao Technology and the intelligent customer service company Xiaoduo Technology also shared their case experiences of connecting to Doubao Big Model to achieve business growth. Among them, the average daily call volume of Doubao Big Model of Yupao Technology has exceeded 100 million tokens, and the application scenarios cover job requirement identification, job type identification, job search intention identification, recommendation system similarity identification, etc.

In the future, Volcano Engine will continue to explore the practical application of big models in thousands of industries, continue to accumulate the practical experience of ByteDance's internal and external customers, and help enterprises implement AI transformation through Doubao Big Model and Volcano Ark's full-stack AI services, release growth potential, and achieve commercial value to achieve "cash" growth. Let all walks of life use Doubao Big Model to make money.

3. The "Bean Bag Family" has a strong lineup and each one can fight:

The Doubao model family was officially released in May and provides a model family with multimodal capabilities. It mainly includes nine models, including the general model pro, general model lite, speech recognition model, speech synthesis model, and text and image model. Each model is gifted and super powerful.

This time, the Doubao model family has upgraded the Wenshengtu model and speech model. The upgraded Doubao Wenshengtu model has the ability to better understand "Chinese" and the user's expression more accurately, and can generate high-aesthetic pictures with consistent text and pictures. After the upgrade, the Doubao speech synthesis model has realized intelligent recognition of text emotions and dynamic adjustment of speech speed and tone to make it more emotional; the sound reproduction model can reproduce human voice with high fidelity in just 5 seconds, and supports a variety of minority languages.

It is worth mentioning that the Doubao model family has officially announced a new member - Doubao · Tusheng picture model. Built on the Wensheng picture model, Doubao · Tusheng picture model can achieve a high degree of restoration of character features and generate pictures that are more like your own. More than 50 styles and scenes can be freely switched. This model capability has been implemented in ByteDance apps such as Douyin, Jianying, Doubao, and Xinghui, and has served corporate customers such as Samsung and Nubia through the Volcano Engine. The current average daily number of Tusheng pictures has reached tens of millions.

Samsung Galaxy AI's new Smart Portrait feature uses the Doubaotu model to create single-image portraits, which enhances the user's photography experience with stronger image processing capabilities, allowing users to efficiently process the photos they take in a more personalized way, adding practicality and fun to the photos. The Volcano Engine AI solution has been continuously refined by products with hundreds of millions of DAUs and has outstanding capabilities in AI portraits. Samsung users only need to upload a single photo to transform it into a new picture in a variety of different styles, such as business, 3D cartoon, cyberpunk, etc., to achieve personalized application of pictures.

Since the Doubao APP was launched one year ago, it has more than 26 million monthly active users. These practical applications of "killing monsters and upgrading" have made the Doubao model more and more powerful. More and more users use Doubao for work, life and study.

4. The ByteDance large-model work ecosystem has been established:

In addition to good results and low prices, Doubao Big Model also provides a limit of Tokens processed per minute that is several times higher than that of the same-tier models, giving the model service a stronger carrying capacity. According to Sun Fan, the algorithm architect of Volcano Engine Big Model Service, Doubao Universal Model Pro provides customers with the highest initial TPM (Tokens per minute) and RPM (Requests per minute) standards in the industry, which helps companies implement their business in high-concurrency scenarios. "We hope to use our solid technical strength to give customers better choices, help companies relieve cost burdens, and allow customers to try and iterate more actively and boldly, so that big model applications can move forward in big strides." Sun Fan said.

Volcano Engine has also upgraded the same plug-in service as Toutiao and Douyin, adding new plug-ins such as web page parsing and calculator, further expanding the boundaries of model capabilities and supporting diverse application needs of enterprises. It also brought the latest developments in multiple products and cloud infrastructure such as Button Professional Edition and HiAgent Platform.

Tan Dai, president of Volcano Engine, said in an interview that the ultra-low pricing of large models comes from confidence in technology, and Volcano Engine can optimize the inference cost of large models through technical means. At Volcano Engine's 2024 AI Innovation Tour, the trend of "application is king" in the current era of large models was confirmed. In the ecological universe of Doubao's large models, we can board the Volcano Ark and join hands with Doubao to "embrace all rivers and seas" and go to the sea of ​​stars that belongs to large models!

Cover News reporter Li Qi