Under the domestic "model" war, who is the strongest, the world's leading IT market research and consulting company IDC's latest "AI Large Model Technical Capability Evaluation Report, 2023" gives the answer. The report shows that Baidu Wenxin Model 3.5 scored 7 out of 12 indicators, ranking first in comprehensive score, algorithm model, and industry coverage. The three absolute firsts reflect the depth of basic technology and the breadth of industrial application coverage of Baidu Wenxin Model.
[IDC "AI Large-scale Model Technical Capability Assessment Report, 2023": Baidu has 7 full scores and the comprehensive score is the first]
The IDC evaluation report examines more than 10 indicators of the large model around the three dimensions of product technology, service ecology, and industry application. Among them, "algorithm model" and "industry coverage" have become two extremely important indicators for measuring the capabilities of large models.
At present, the large model is in a stage of rapid development, and product technology capabilities and industry application capabilities are particularly important.
In product technical capabilities, the "algorithm model" dimension is the most important, the core element of large model capabilities, and the root of determining the application effect of large models. Only through the breakthrough of algorithm model technology and the realization of a large model base with general effect advantages can it support wider industry coverage, enable all walks of life to fully enjoy the dividends brought by technological breakthroughs, and solve the dilemma of high threshold for AI implementation.
In terms of industry application capabilities, the breadth of application coverage is currently the most concerned indicator for large-scale model manufacturers. "Industry coverage" reflects the strength of the large model in industrial implementation through the number of enterprise-level customers and the number of landing industries. It is a comprehensive manifestation of the general leading effect of the large model and the ability to combine industries.
The two core indicators of "algorithm model" and "industry coverage" are inherently related. The breadth of industry coverage is a concentrated expression of the general leadership of the algorithm model. It will also provide a steady stream of positive feedback for the continuous improvement of the algorithm model capabilities and form a flywheel for continuous iterative improvement.
In this evaluation, the two indicators of Baidu Wenxin large-scale model obtained the only full score among many manufacturers, fully reflecting the most advanced technology of Wenxin large-scale model products and the most extensive and in-depth industry applications.
The industry's first large-scale model evaluation framework, Baidu Wenxin's comprehensive score No. 1, the only algorithm full score
This is the first time that IDC has proposed an evaluation framework for AI large-scale model technology capabilities. 14 domestic mainstream large-scale model manufacturers, including Baidu, Ali, Tencent, Huawei, HKUST Xunfei, 360, and SenseTime, participated in this evaluation. The results show that the overall competitiveness of Baidu's AI large model is at the leading level, and it is the technological breakthrough and application leader of this large model. Baidu Wenxin has obvious advantages in model capabilities, tool platforms, ecological layout, and industry coverage, and has entered the stage of commercialization and landing exploration ahead of schedule.
[IDC "AI Large-scale Model Technical Capability Assessment Report, 2023": Baidu won the only perfect score for algorithm model and industry coverage]
Wenxin large model obtained the only full score in the algorithm model dimension in this IDC evaluation, which fully reflects Baidu's leading edge in the core technology of large models. Baidu has been deeply involved in the research and development of pre-training models since 2019, and has successively released the knowledge-enhanced Wenxin series of models. Not long ago, Baidu officially released version 3.5 of Wenxin Large Model, which further made innovations in multiple core technologies such as basic model, knowledge enhancement, and retrieval enhancement. The new version has significantly improved various capabilities.
A number of public evaluations show that Wenxin Yiyan, supported by the Wenxin large model version 3.5, has outstanding Chinese ability, even surpassing GPT-4; its comprehensive ability surpasses ChatGPT in the evaluation, and is far ahead of other large models.
Wenxin's large model achieved "No. 1" thanks to the advantages of Baidu's four-layer technology stack of "chip-framework-model-application", the core feature of knowledge enhancement, and the prosperous large-scale model ecology.
According to reports, Baidu has a self-developed deep learning platform Flying Paddle, which strongly supports the efficient training and reasoning of large models. The collaborative optimization of Flying Paddle and Wenxin has improved the model effect of the latest version of Wenxin Large Model 3.5 by 50%, increased the training speed by 2 times, and increased the inference speed by 30 times. As one of the core features of the Wenxin model, knowledge enhancement has achieved higher efficiency, better effects, and stronger interpretability.
In terms of large-scale model ecology, Baidu Wenxin has formed an ecological system integrating enterprise, education, and community. The latest data shows that Baidu has a developer base of more than 7.5 million and an ecological base of 200,000 enterprises. It has carried out large-scale talent training, enterprise empowerment, and developer operations at multiple levels. Baidu also set up a 1 billion venture capital fund to encourage large-scale model creativity and prosper the large-scale model ecology.
Wenxin has the largest industrial application scale in China, and its industry coverage has been awarded the only full score
The AI large model has developed from a competition of parameters to a competition of applications, and has entered the stage of large-scale replicable industrial implementation. Baidu Wenxin's large-scale model originated from industrial practice and serves industrial practice. For the first time in the industry, it proposed the idea of implementing large-scale industrial models. It cooperated with State Grid, Shanghai Pudong Development Bank, Taikang, Geely, Harbin, Shenzhen Gas, TCL, Shanghai Dictionary Publishing House and other enterprises to jointly release 11 large-scale industrial models. It is the earliest manufacturer in the industry to promote large-scale industrial models extensively and deeply.
[Baidu Wenxin large model panorama]
The IDC evaluation results show that Baidu Wenxin's large model has achieved the only full score in industry coverage, and has achieved extensive business layout and landing scenario exploration in the fields of energy, finance, education, and medical care.
According to reports, the Wenxin large model already has the largest industrial application scale in China, and currently 150,000 companies have applied for access to the Wenxin Yiyan test. Baidu Smart Cloud and more than 300 ecological partners have achieved quite good test results in more than 400 scenarios.
Taking energy and electricity as an example, in the world's largest public utility company, State Grid Corporation of China, for professional scenarios of complex power grids, based on the Baidu Wenxin large model, Baidu and the Institute of Intelligence jointly train the large model of the power industry, and carry out pilot verification in actual business scenarios such as power grid equipment and customer service, which can significantly enhance the level of refinement, automation, and intelligence of power grid operations. Baidu and Shenzhen Gas jointly released a large-scale model of the gas industry to solve problems such as complex operating scenarios and difficult identification of safety risks for gas companies.
In the future, all enterprises will strongly rely on big models, and all products will be developed based on big models. Baidu Wenxin's big model will continue to give full play to the basic technical advantages of the algorithm model, helping Chinese enterprises in all industries to internalize the big model as their own productivity tool, embark on the fast track of intelligent transformation and upgrading, and build a strong global competitiveness.