Baidu Intelligent Cloud Qianfan large model platform has been upgraded again: 5 large models and 55 new tool components have been released!

Serves 80,000 enterprise users, has helped users fine-tune 13,000 large models, and helped users develop 160,000 large model applications. Since December 2023, Baidu Intelligent Cloud Qianfan Large Model Platform API daily calls increased by 97% month-on-month... From the "pioneer" of the domestic large model platform a year ago to the large model "super factory" today, Baidu Smart Cloud Qianfan large model platform firmly occupies the domestic large model market He was in the lead, but his running pace did not stop.

On March 21, Baidu Intelligent Cloud held a Qianfan product launch conference in Beijing Shougang Park. Baidu Intelligent Cloud announced during the conference:

1. Join hands with Beijing’s Shijingshan District to build the country’s first Baidu Intelligent Cloud Qianfan large-scale model industrial innovation base to boost the regional industry’s take-off;

2. To meet the core demands of enterprises for “cost-effectiveness”, Qianfan ModelBuilder large model services have been fully upgraded, with 3 lightweight large models and 2 vertical scene large models newly released;

3 , significantly lowering the threshold for AI native application development, and Qianfan AppBuilder component capabilities have been fully upgraded.

(Baidu Intelligent Cloud Qianfan Large Model Industry (Beijing) Innovation Base Starts Construction)

During the event, Cao Shihui, a member of the Shijingshan District Government Party Committee and Deputy District Mayor, Cui Mingming, Deputy Director of the Zhongguancun Shijingshan Park Management Committee, Xie Guangjun, Baidu Vice President, and Shi Qinghua, Baidu Vice President, launched the country's first The launching ceremony of Baidu Intelligent Cloud Ganfan Large Model Industry (Beijing) Innovation Base.

The innovation base is committed to promoting the deep integration of large model technology and industrial innovation. The two parties will focus on improving the supply of computing power, optimizing model algorithms, and promoting data openness , create demonstration scenes, deepen talent introduction and education, etc., and the government and enterprises will work together to create a new highland for the artificial intelligence industry. Deputy District Governor Cao Shihui said that Shijingshan District will work hand in hand with Baidu to focus on the research and development and innovative application of artificial intelligence and large model technology, build and improve the AI ​​industrial ecosystem, provide support for regional digital transformation and industrial intelligent upgrading, and provide support for new quality productivity. Cultivation and development are infused with surging momentum.

(Baidu Vice President Xie Guangjun)

2023 , large models have shown explosive growth on a global scale, and the "Battle of 100 Models" has started in China. Countless large technology manufacturers and scientific research institutes have focused on the "ontology" of large models and are "involving" crazily.

Xie Guangjun, Vice President of Baidu, said that large model technology has developed rapidly in the past year. With its gradual implementation in thousands of industries, 2024 will be the first year for the explosion of domestic large model industry applications. In response to the four major challenges that enterprises are most concerned about, including large model implementation scenarios, usage costs, application development, and application effects, Baidu Intelligent Cloud Qianfan has provided the latest "problem-solving ideas" in two aspects: large models and AI native application development.

Baidu Intelligent Cloud Qianfan large model platform releases "3 2" new model package: improving the "cost-effectiveness" of enterprise application large models ”

The effect of large models is the unremitting pursuit of the “technical school”, while economic benefits are the ultimate goal of the “market school”. In communicating with customers in many industries, Xie Guangjun found that except for a few large customers who pursue the ultimate effect of large models, more companies and institutions often need to comprehensively consider the use effect, performance and cost of large models, that is, "valency" Compare". This time, the Qianfan platform’s model matrix has undergone a series of upgrades in response to the company’s core appeal of “cost-effectiveness”.

(Baidu Intelligent Cloud Qianfan Large Model Platform Model Matrix)

Wenxin large model ERNIE 3.5 is currently one of the most popular basic large models on Baidu Intelligent Cloud Qianfan large model platform. For users' common dialogue scenarios, ERNIE 3.5 has enhanced its capabilities in three aspects: instruction following, context learning and logical reasoning capabilities. The upgraded ERNIE 3.5 has significantly improved application performance by 24%, 27% and 22% in three major enterprise application scenarios, such as copywriting creation, information extraction and tool invocation.

#Secondly, compared to large models with very large parameters, lightweight large models have smaller parameters, making it easier for customers to target specific usage scenarios. Fine-tuning the model makes it easier to achieve expected results and save more costs. At this conference, Baidu Intelligent Cloud released three lightweight models including ERNIE Speed, ERNIE Lite, and ERNIE Tiny, with parameters ranging from large to small, to help customers achieve "reduce volume without reducing efficiency" and save unnecessary costs. invest.

Specifically, ERNIE Speed, as the "big one" among the three lightweight large models, has up to 128k context in inference scenarios. Length, when dealing with tasks such as knowledge question and answer, can better handle context dependencies and generate more coherent and accurate predictions or answers. At the same time, ERNIE Speed ​​can be used as a base model for fine-tuning for specific scenarios. The model effect can equal or even exceed the flagship large-parameter large-scale model, and the cost-effectiveness is greatly improved.

Compared with ERNIE Speed, ERNIE Lite has smaller parameters and is more suitable for processing inference on AI accelerator cards with low computing power. Tasks, while taking into account the model effect and inference performance, greatly reduce the cost of customer landing application. As an upgraded version of the ERNIE-Bot-turbo model, ERNIE Lite's application effects in scenarios such as sentiment analysis, multi-task learning, and natural reasoning have improved by 20%. The cost of inference calls has dropped significantly by 53%!

ERNIE Tiny, which has the smallest number of parameters among the three lightweight models, provides customers with the best choice for ultimate low cost and low latency. In high-concurrency, low-latency application scenarios such as retrieval, recommendation, and intent recognition, ERNIE Tiny's excellent performance shows outstanding performance. In a certain dialogue recommendation business scenario, the fine-tuned ERNIE Tiny increased the number of dialogue rounds by 3.5% and reduced the cost by 32% compared with ERNIE 3.5 in stimulating search engine recommended words.

#In addition, enterprises have higher requirements for the effects of large models in character playing and external tool calls in practical applications. This Qianfan large model platform ModelBuilder also launched two vertical scene large models, ERNIE Character and ERNIE Functions, based on in-depth insights into enterprise scenarios and Baidu's own business best practices, to adapt to customers' role-playing application scenarios ( Such as game NPC, customer service dialogue, etc.) and tool calling scenarios (using external tools in dialogue, calling business functions, etc.).

#In corporate practice, a smart hardware manufacturer created an intelligent assistant based on ERNIE Character. After applying this model, it improved character consistency, motivation and The effect of improving users' desire to chat has been significantly improved. A travel and travel APP uses ERNIE Functions to create an intelligent customer service assistant, with an accuracy of 85% in executing multiple function calls such as booking tickets and checking flight status.

##Qianfan AppBuilder fully upgraded: significantly lowering the threshold for AI native application development

QianfanAppBuilder, as an industrial-grade AI native application development platform, is an important part of Qianfan. The bottom layer of AppBuilder is composed of large model components based on Baidu's years of technology and practical experience, basic components of AI capability components, and a series of advanced components that are in-depth tuned and constructed for typical application scenarios. Basic components and advanced components jointly support the Agent. On the one hand, more complex business logic can be implemented through workflow orchestration. On the other hand, the Agent also has strong autonomous task planning capabilities and can understand user intentions and automatically plan execution paths to achieve multi-tool integration. Automated orchestration and execution. These capabilities are provided through two development methods, zero-code state and code state, to better match the needs of different developers.

After the development is completed, the application can be distributed and integrated through multiple channels. AppBuilder supports one-click distribution of applications to WeChat customer service, WeChat official accounts, and the Web. /H5 and Baidu Lingjing Matrix and other mainstream channels. Based on the Baidu Spiritual Realm Matrix, applications can be distributed and mounted in mainstream scenarios such as Baidu search and Baidu information flow. Once the application is truly developed, it will directly reach users and open up the entire process from AI native application creation to development to distribution.

（Baidu Intelligent Cloud Qianfan AppBuilder）

The upgraded AppBuilder has as many as 55 open tool components. It includes large model components and AI capability components based on Baidu's years of technology accumulation and its own business accumulation, as well as Baidu-specific business components such as search, and third-party API tools for multiple scenarios. In addition, it also provides RAG (knowledge retrieval question and answer), Advanced capability components such as GBI (Generative Data Analysis) that are in-depth tuned according to typical application scenarios.

Developing AI native applications is inseparable from cloud infrastructure. The basic components released this time also include Baidu Smart Cloud’s newly launched vector database VDB 1.0. Vector database is an indispensable core component of knowledge base for enterprises. It provides a powerful solution to the performance bottlenecks, maintenance challenges and scale limitations encountered by traditional knowledge base question and answer systems. The newly released Baidu vector database VDB 1.0 not only integrates comprehensive operation and maintenance control and security protection capabilities, but is also compatible with mainstream ecosystems such as Qianfan and LangChain, which can help enterprises easily manage tens of millions of document knowledge and support up to 100 Billion vector storage scale and millisecond-level vector retrieval speed. At the same time, compared with similar open source products, the performance of VDB 1.0 is improved by up to 10 times.

On top of the components, the Agent (intelligent) application framework launched by Qianfan AppBuilder has precise independent task planning capabilities and is suitable for a variety of applications. The tool's automatic orchestration accuracy exceeds 90%, and this number is still improving. AppBuilder also supports developers to access custom tools and combine automatic orchestration with manual orchestration to customize the needs of more complex scenario applications. The code interpreter capability within the Agent framework has also significantly improved performance by 40% in this upgrade, and the acceptance rate of generated results in complex data analysis scenarios is as high as 95%, making it easy to cope with various data analysis and information processing scenarios. .

In addition, AppBuilder’s code-based development tools are also a new addition. AppBuilder SDK releases Agent API this time, allowing developers to easily integrate Agent into their business systems. At the same time, AppBuilder SDK provides a wealth of application examples for mainstream AI native application scenarios. It is currently open source on Github and supports various applications. Flexible orchestration with free call of components helps developers realize secondary development and convenient integration of applications.

At the conference, it was also demonstrated how to build an "English composition helper" Agent (intelligent (body) application, just enter the application name or the application function you want to develop in AppBuilder, and the platform can automatically generate the application. By simply adjusting the role instructions and adding the required tool components, you can quickly generate an English composition correction assistant. You can use it directly after publishing, and complete application creation and distribution in three steps.

(Create an English composition correction assistant in 1 minute)

In the view of Baidu Intelligent Cloud, with the continuous evolution and breakthrough of large model technology, the adaptation of engineering practice and user needs is becoming increasingly important. Only by going deep into the scenario, discovering and responding to the real needs of customers is the best way to release the dividends of innovative technology.

The above is the detailed content of Baidu Intelligent Cloud Qianfan large model platform has been upgraded again: 5 large models and 55 new tool components have been released!. For more information, please follow other related articles on the PHP Chinese website!