Introduction: To solve the challenge of 'application difficulty' of large models, Baidu flexed its muscles in its Q2 financial report.

Ludwig Beethoven, who had philosophical musings on AI, famously stated, 'The meaning is in the use.'

Setting aside its potential philosophical implications, this statement aptly captures the essence of AI in a practical context: it must be user-friendly.

Since 2020, with the rise of large models and rapid advancements in natural language processing technology, AI technology has entered a period of accumulated strength and potential outburst.

After the technological explosion, the 'application difficulty' of large models gradually emerged. 'How to effectively use AI' has become a new challenge for enterprises when deploying AI on a large scale.

Judging from capital flows and corporate financial reports, this challenge is being addressed.

Taking the Wenxin Large Model as an example, Baidu's (BIDU.US; 09888.HK) Q2 2024 financial report conference call revealed that the daily call volume of the Wenxin Large Model exceeded 600 million times, with a more than tenfold increase in half a year. The daily processing of over 1 trillion tokens is also the highest in China. Meanwhile, the revenue from intelligent cloud services reached RMB 5.1 billion in this quarter, an increase of 14% year-on-year. AI-driven cloud revenue also increased by 9% year-on-year, higher than the 6.9% in the previous quarter.

Whether empowering external industries or supporting its own business, Baidu's large model has delivered impressive results at the application level, demonstrating its strength.

01 Wenxin Large Model, Racing Forward

Amidst the 'Hundreds of Models War,' the number of large models has increased, but their practical applications remain a weak spot.

Giants have entered the fray, and small and medium-sized tech companies also want a piece of the pie. However, compared to their sheer numbers, their shortcomings in application stand out. Many large models are shelved, gathering dust, unable to enter the vast application market.

Some research institutions suggest that technical indicators alone should not be the sole criterion for judging the success of a large model. Instead, more attention should be paid to tangible business metrics.

Undoubtedly, the direct business metric for measuring whether a large model is accepted by the market is its daily call volume.

Unlike some companies that are tight-lipped about their call data, Baidu openly disclosed the daily call volume of the Wenxin Large Model during its Q2 financial report conference call:

During the quarter, the daily call volume of the Wenxin Large Model exceeded 600 million times, with a daily usage of over 1 trillion tokens, both the highest in China.

Compared to other large models with daily token usage ranging from tens of billions (MiniMax and Zhipu AI) to hundreds of billions (Doubao Large Model), it is no exaggeration to say that Baidu has made the fastest progress in applying large models among domestic manufacturers.

While volume is one aspect, the growth rate is even more noteworthy. According to Baidu's Q1 financial report, the daily API call volume of the Wenxin Large Model was approximately 200 million times, compared to just 50 million times in December last year.

In half a year, there has been an increase of over tenfold.

The rapidly growing call volume indicates that the large model is 'more user-friendly,' offering faster speeds, better experiences, and lower prices.

The core advantages of the Wenxin Large Model 4.0 Turbo lie in speed and effectiveness, thanks to Baidu's continuous innovations in core technologies such as data, basic models, alignment techniques, prompts, knowledge enhancement, retrieval enhancement, and dialogue enhancement.

Advantages of the Wenxin Large Model, Source: Wenxin Large Model Official Website

In terms of pricing, the Wenxin Large Model 4.0 Turbo is fully open to enterprises, with a pricing model of RMB 0.03 per thousand tokens for input and RMB 0.06 per thousand tokens for output.

This move signals another direction in Baidu's recent layout for large model applications: price reductions and free strategies.

Baidu Vice President Xie Guangjun stated that compared to the Wenxin Large Model 4.0 General Edition, the Wenxin Large Model 4.0 Turbo offers a 70% price reduction under a 3:1 input-to-output ratio.

Previously, the flagship models ERNIE 4.0 and ERNIE 3.5 of Wenxin had already announced significant price reductions, and the pre-installed services of the ERNIE-Speed, ERNIE-Lite, and ERNIE-Tiny series models have been continuously made available to customers for free.

As the large model market matures, price competition is inevitable. For small and medium-sized enterprises with limited budgets, this is a cost-effective measure. For large model vendors, it is also a 'gatekeeper' to sift out the genuine and test their capabilities.

Beyond iteration, price reductions, and inclusiveness, the enormous call volume of the Wenxin Large Model is undoubtedly the result of 'voting with feet.' Under natural business laws, users will actively choose products that are more user-friendly and affordable, and the resulting economies of scale will, in turn, benefit the large model, creating a virtuous cycle where 'the more people use it, the better it becomes.'

02 Large Models 'Blood Transfusion' for Cloud Business

With the Wenxin Large Model leading the way, an increasing number of startups and ecosystem partners are gearing up to participate in this development wave.

As a result, MaaS (Model as a Service) has become popular, and 'model factories' have emerged. As cloud platforms providing data, models, and services, they have found a new revenue path.

Baidu, which has long been involved in cloud services, has validated this high-value revenue path in its Q2 financial report:

Baidu's intelligent cloud revenue reached RMB 5.1 billion, an increase of 14% year-on-year, and continued to achieve profitability (Non-GAAP). Meanwhile, AI-driven revenue accounted for an even higher proportion of 9%, up from 6.9% in the previous quarter. According to an IDC report, Baidu Intelligent Cloud ranked first in China's large model platform market in 2023, with a market share of 19.9%.

This growth is attributed to Baidu Cloud's deep involvement in the MaaS service platform.

To facilitate the practical application of large models, Baidu launched the Qianfan Large Model Platform, the world's first one-stop enterprise-grade large model platform, supporting fine-tuning of six Wenxin Large Model variants, including the flagship ERNIE 4.0 Turbo. To date, the Qianfan platform has fine-tuned 21,000 models, serving the core business scenarios of over a thousand enterprises.

On the platform, two development tools are provided to clients: ModelBuilder for model customization and AppBuilder for AI-native application development.

In Q2, the usage of both tools increased significantly. Qianfan's confidence stems from three factors: firstly, access to over 70 mainstream foundational and industry-specific large model resources from China and abroad, including the Wenxin Large Model; secondly, a highly user-friendly toolchain that economically and easily assists users in model training, fine-tuning, deployment, and AI application development; and thirdly, compatibility with various industries. In May this year, Baidu Intelligent Cloud launched the 'Qianfan Industry Enhanced Edition,' integrating the comprehension, generation, logic, and memory capabilities of general large models with the professional capabilities of industry-specific models, supported by systematic tools and components, covering the entire process from resource management and large model services to application development.

During the quarter, the Qianfan platform delivered tangible value to industries such as government, electricity, and education, enabling rural areas to benefit from large models. From January to May 2024, nearly 260 domestic large model project bid announcements were released, with Baidu winning the most projects (10), covering the most industries (8), and securing the highest total bid amount (over RMB 64 million), ranking first in all three categories among mainstream large model vendors.

03 The Eruption of Intelligent Agents

However, the above is not the final form of Baidu's AI layout.

Looking ahead, Baidu's focus lies in intelligent agents.

During the 2024 World Artificial Intelligence Conference, Baidu's founder, chairman, and CEO Robin Li proposed that intelligent agents represent the future trend of the AI era. 'Intelligent agents are erupting, but their current base is still relatively small, so the perception may not be as strong.'

And Baidu has seized the initiative in both strategy and performance. The Q2 financial report conference call revealed a substantial increase in the distribution of intelligent agents within the Baidu ecosystem, with over 8 million daily distributions in July, double that of May, indicating the gradual formation of a Baidu intelligent agent ecosystem.

What are intelligent agents? As early as the 1940s, Marvin Minsky proposed the concept of agents in 'The Society of Mind,' suggesting that tiny components of intelligence could form true intelligence in a very special way.

In the AI realm, intelligent agents refer to programs, systems, or applications with perception, action, autonomous decision-making, and goal-oriented capabilities.

In Li's view, intelligent agents are nearly universal applications based on large models. Due to their low threshold, users can create effective intelligent agents without even needing to program.

The eruption of intelligent agents implies that in the future, industries such as finance, healthcare, education, agriculture, manufacturing, and transportation will create various intelligent agents based on their unique scenarios, experiences, rules, and data, forming a vast ecosystem of millions.

Moreover, there is another consequence: search, as the primary gateway for intelligent agent distribution, will become increasingly vital.

On the one hand, search is a natural scenario for intelligent agent deployment and distribution, connecting a vast user base and serving as an excellent distribution pool. Baidu Search launched the 'AI Assistant' channel, aiming to provide users with matching intelligent agents by recognizing their search intent. In Q2, 18% of search results were generated by AI, providing users with more accurate and direct answers.

For enterprises and developers, Baidu can also leverage its search ecosystem to distribute and monetize their intelligent agents, helping clients enhance sales conversions and improve advertising efficiency.

On the other hand, intelligent agents are reshaping AI search, fully unleashing Baidu's potential. Since the eruption of large models, search engines worldwide, including Google, Microsoft, and Meta, have been exploring ways to enhance search capabilities through large models.

In China, Baidu stands unparalleled in the search field. With a daily request volume of billions, a vast user base, immense traffic, and intelligent recommendations, Baidu is uniquely positioned to explore the integration of 'intelligent agents and search.'

In the present, Baidu has emerged as a bridgehead in overcoming the 'application difficulty' of large models, with the daily call volume of large models racing forward and the rapid deployment of the MaaS platform.

Looking into the future, the eruption of intelligent agents is unstoppable, and the burden of 'intelligent agents + search' will inevitably fall on Baidu's shoulders.

The AI era calls for true capabilities. At least based on the Q2 financial report, Baidu, basking in the dawn of general AI, has lived up to expectations.