Baidu has once again revamped its AI organizational framework.

The latest move involves the creation of the Baidu Model Committee (BMC). Previously established, the Basic Model Unit (BMU) and Applied Model Unit (AMU) will now fall under the purview of the BMC; pertinent departments will also maintain their reporting line to Baidu CEO Robin Li.

According to publicly available information, the BMC comprises young researchers with a profound grasp of large-scale models and will oversee the R&D endeavors for both foundational and applied models at Baidu.

The BMC introduces a coordinating layer above the BMU and AMU, marking the formalization of this structure following the establishment of the new model R&D departments last year.

The underlying implication of this reorganization is that Robin Li is carving out a dedicated AI domain within Baidu.

This domain is youthful, closer to the forefront of large model R&D, and characterized by a streamlined decision-making process. It aims to circumvent the persistent challenges Baidu has encountered over the years: missing out on opportunities despite their abundance.

Baidu: A Visionary Company

Baidu has consistently been at the forefront of technological foresight.

Among Chinese internet firms, Baidu was one of the earliest to systematically invest in AI. Whether it's smart speakers, autonomous driving, or later, large models, Baidu didn't jump on the bandwagon only after witnessing others' success. Often, Robin Li's technological trend predictions were even ahead of their time.

Yet, this very foresight is where Baidu's predicament lies.

If a company consistently fails to recognize trends, its challenges are straightforward. However, if a company repeatedly identifies trends early but consistently lags when it comes to delivering results, the issues become far more intricate.

In recent years, Baidu has been frequently described as 'early to rise but late to arrive.' This phrase has gained popularity due to its accurate depiction of the core contradiction in Baidu's AI journey: a strong sense of direction but a lack of compelling results.

Robin Li emphasized AI early on, discussed intelligent agents, and consistently aimed to shift the AI competition from mere model capabilities to applications and productivity. However, the crux lies in the fact that Baidu's delivered products and models often failed to substantiate his predictions.

The most notable example is the debate over open-source versus closed-source models in the past two years.

Robin Li sparked widespread discussion with his assertion that open-source models had no future and that closed-source models were the way forward. However, from a global perspective on large model commercialization, this assertion is not entirely unfounded.

The mainstream commercial model paths, exemplified by companies like OpenAI, Anthropic, and Google, essentially validate this point. Closed-source models can establish barriers, pricing systems, and generate commercial revenue.

Today, domestic large model companies' open-source and commercialization strategies are not purely open-source. Alibaba adjusted its open-source strategy, and Doubao is preparing to implement a paid plan.

If Baidu's ERNIE Large Model were clearly leading in terms of capabilities, experience, and industry reputation, Robin Li's assertion about closed-source models would seem logical. However, when the external perception is that ERNIE lacks sufficient dominance, a potentially valid commercial prediction becomes a target for ridicule.

This suggests that Robin Li's issue may not be a complete misjudgment of direction but rather an overestimation of Baidu's own model capabilities.

Where does this overestimation stem from? Largely from the information input within the organization.

A CEO's assessment of a company's core capabilities cannot rely solely on personal product testing in the office. It depends on reports, internal evaluations, and the layer-by-layer transmission of R&D progress, model metrics, product feedback, and competitive judgments.

If this information is filtered, packaged, and overly optimized by organizational inertia, what ultimately reaches the CEO is not the real-world picture of Baidu.

This is the backdrop against which the BMC, as well as the BMU and AMU, must be understood.

Carving Out a Domain, Bypassing the Old System

The most noteworthy aspect of the BMC is its reporting relationships and the involvement of 'young researchers.'

These aspects are more significant than the 'committee' itself. They indicate that Baidu is not merely seeking higher-level management coordination but is attempting to bring frontline technical personnel who truly understand large models closer to the decision-making center.

This carries at least three layers of meaning.

First, Baidu hopes to reduce the filtering of technical judgments by traditional management chains.

Large model R&D is not a field suitable for layer-by-layer retelling. Model roadmaps, training data, compute allocation, post-training strategies, and application feedback all evolve rapidly and heavily rely on frontline judgments. If decision-makers primarily hear summaries processed through multiple layers, they can easily overlook crucial technical details.

Having young, model-savvy researchers in the BMC means Baidu aims to elevate frontline R&D judgments to a higher level. It's not merely a gesture of youthfulness but an attempt to alter how information reaches Robin Li.

Second, this approach aligns with the organizational ethos of AI startups.

In recent years, many of the world's strongest AI companies have not adhered to the typical organizational forms of traditional giants. Their core technological judgments often come from frontline teams very close to model training and product iteration. The distances between research, engineering, product, and commercialization are short, allowing for rapid course corrections and making it difficult to conceal deficiencies for long.

Foreign giants largely lack a presence in model development, influenced by both capability limitations and active choices. Google, the only one to achieve significant results, relies on DeepMind as a relatively independent R&D facility.

As a mature large company, Baidu cannot fully replicate a startup's structure. However, the BMC at least moves in that direction: ensuring model R&D is no longer entirely buried within existing organizational hierarchies but forms a shorter, more direct pathway.

Third, this is a correction to Baidu's past internal information chains.

If Baidu's past issue was 'early to rise but late to arrive,' it must reflect not just on the strategic level but also on the execution, feedback, and judgment levels. Robin Li needs to know the true model capabilities, product experiences, and competitive gaps. The more these pieces of information pass through layer-by-layer reporting, the more likely they are to become distorted.

The BMC's establishment essentially shortens this distance.

A more subtle signal can also be discerned here: Baidu's traditional AI technical system has long been associated with Wang Haifeng. However, the BMU, AMU, and now the BMC all report directly to Robin Li's large model line. At the very least, this indicates that Robin Li hopes to separate large model R&D from Baidu's existing technical management system, forming a shorter, more direct decision-making chain.

BMU and AMU Should Not Be Separated

The BMU focuses on general-purpose foundational models, while the AMU develops more application-specific, specialized models and capabilities. On the surface, one leans toward the base, and the other toward scenarios. However, in the large model era, the two are inherently interdependent.

Whether a programming-related application can succeed depends not just on how the application layer is designed but also on the starting point of the foundational model's coding capabilities. This starting point, in turn, influences training data weights, evaluation systems, and model roadmaps.

If the foundational model team only pursues generic metrics and the application model team only supplements capabilities at the backend, a disconnect will emerge. The BMC's value lies precisely here: it attempts to place foundational model R&D, applied model R&D, and business implementation within the same coordinating mechanism.

Thus, the BMC, BMU, and AMU resemble an internal entrepreneurship mechanism.

Robin Li is carving out an AI-specific domain within Baidu because he may have realized that the existing organizational system is unlikely to naturally produce the large model results he desires. Since the current system cannot be relied upon, he is experimenting with a form of internal entrepreneurship—a kind of 'going solo.'

BMC Reorganizes the Team, DAA Defines the Rules

At Create 2026, Robin Li introduced another concept: DAA, or Daily Active Agents.

Robin Li's point is that in the AI era, we should not just focus on how many tokens have been consumed or by how many people but rather on how many intelligent agents are completing tasks on behalf of humans. In the mobile internet era, DAU (Daily Active Users) was a core metric. However, in the age of intelligent agents, the agents themselves are the participants.

This direction is insightful.

If the value of AI lies not just in answering questions but in executing tasks on behalf of humans, then the metric should indeed shift from 'how many times humans accessed it' to 'how much work machines accomplished.' In this sense, DAA represents Baidu's attempt to propose a new scoring system for the intelligent agent era.

However, DAA also has its issues.

An agent is essentially a program endowed with a certain degree of intelligence. As long as one is willing to split, deploy, and run them, the number of agents can be rapidly increased. Whether an agent truly creates value cannot be judged solely by its 'activity.'

On the contrary, low-quality agents may even create noise.

AI comments on social media provide an intuitive example. They superficially increase interaction but do not necessarily offer new information, experiences, or judgments. Users read comments to gain insights they couldn't arrive at independently.

If comments are merely generic phrases generated effortlessly by large models, they add to the cognitive burden. If I want to see AI replies, I might as well ask the AI directly. Moreover, the models running agents are probably not the top-tier ones. If I ask a question, I can choose between Opus 4.7 or GPT-5.5.

Robin Li said that tokens represent costs, not outputs, which is partially true. Token consumption alone does not indicate whether a task has been completed. However, conversely, tokens are paid resources and represent real costs.

That enterprises and users are willing to continuously pay for tokens already indicates that they create value in certain areas. Money remains one of the most effective measures, and payment is the strongest form of output recognition.

DAA can be seen, to some extent, as Baidu's attempt to vie for influence in defining the rules of the AI application era.

Viewed alongside the BMC, it becomes even more intriguing.

At Create 2026, Baidu stated externally that AI competition should not focus solely on model leaderboards, parameters, tokens, and chatbot DAUs but rather on whether intelligent agents are truly completing tasks.

This is Baidu's attempt to redefine the external competition rules.

The BMC represents Baidu's internal actions. If future competition truly hinges on whether intelligent agents can consistently deliver, then foundational models, applied models, training roadmaps, and business implementation cannot remain isolated. They must be integrated in a way that ensures resource allocation, efficiency, and information flow.

This is Robin Li reorganizing the internal team to participate in the competition.

However, Baidu is missing one crucial element: native AI application scenarios.

Baidu App, Search, Baidu Netdisk, and Baidu Wenku all have users, data, and scenarios. The issue is that most of these products were formed during the previous internet cycle. Their functions, user mindsets, and business models are already established. AI can transform them, but such transformations mostly enhance existing functionalities.

Miaoda, DuMate, Famou, and Yijing form Baidu's intelligent agent matrix showcased at Create 2026. They complete Baidu's AI application narrative: general-purpose intelligent agents, coding agents, decision-making agents, and digital human agents. However, they still need to prove themselves as stable, high-frequency, and sufficiently monetizable real-world applications.

Organizations can be rebuilt, but application scenarios cannot naturally emerge from organizational structures.

This is the boundary of the BMC and the true challenge Baidu faces next.