Baidu is navigating a pivotal transformation in its business strategy.

Just before unveiling Wenxin Yiyan's deep search function and making it fully free to access, Baidu announced today (February 14) that it will open source its next-generation Wenxin Large Model. According to official news, Baidu will roll out the Wenxin Large Model 4.5 series in the coming months, with official open sourcing commencing on June 30.

Tang Chen revealed that Baidu plans to release multiple models this year and introduce Wenxin Large Model 5.0 in the second half, significantly enhancing the model's multimodal capabilities.

Baidu's Swift Turnaround

Prior to Baidu's announcement of Wenxin Yiyan's free access, OpenAI launched the o3-mini model, comparable to DeepSeek-R1, and made it freely available to users. Baidu explained the sudden free access by stating that the iterative upgrades and costs of the Wenxin Large Model are continuously declining.

However, Baidu did not provide a clear rationale for shifting from a closed-source to an open-source model. This abrupt change surprised many in the industry. Notably, Baidu had previously emphasized 'closed sourcing' and was a staunch opponent of 'open sourcing'.

In 2024, during an internal speech, Robin Li discussed the rationale behind not open sourcing the Wenxin Large Model, citing industry dynamics. He noted that when the Wenxin Large Model was first released, there was considerable debate within Baidu, ultimately leading to the decision to keep it closed source.

Li argued that, in both China and the US, the most advanced foundation models are closed source, and smaller, high-quality models are distilled from these larger ones. Models created by reducing the dimensions of large models are superior, giving closed sourcing advantages in cost and efficiency.

He further believed that closed-source models would continue to lead in capabilities, not just temporarily. Open sourcing large models holds little significance, and closed sourcing is essential for a viable business model. Closed sourcing generates revenue, which in turn attracts computing power and talent.

Robin Li has consistently highlighted the benefits of closed sourcing in public forums. For instance, at the Create 2024 Baidu AI Developer Conference, he bluntly stated that 'open-source models will increasingly fall behind.'

During the 2024 World Artificial Intelligence Conference (WAIC 2024), Li also remarked that open sourcing is akin to an IQ tax. 'Whether it's ChatGPT, Wenxin Yiyan, or other closed-source models, they are undoubtedly more powerful than open-source models and have lower reasoning costs.'

From Baidu's perspective, Robin Li's stance is justified. An unspoken aspect of his argument is that Baidu chose the same closed-source path as OpenAI to ensure market leadership and maximize commercial benefits.

When Baidu opted for the closed-source route, OpenAI, which sparked the large model wave with GPT-2, also turned to closed sourcing. In 2024, OpenAI CEO Sam Altman was even sued by Elon Musk, who mocked him and suggested renaming the company Close AI.

At the time, no one questioned OpenAI's approach or its commercial value. For example, on October 3, 2024, OpenAI announced on its official website that it had received $6.6 billion (approximately RMB 46.3 billion) in funding, making it one of the most highly valued startups in history. The company's post-investment valuation reached $157 billion, approximately RMB 1.1 trillion.

Closed sourcing allows Baidu to monopolize technological advantages and profit by offering API interfaces or enterprise-level services, rather than making codes and technological secrets public and losing control. In essence, Baidu hopes to attract more developers, users, and partners to the Wenxin Large Model ecosystem, where they abide by Baidu's rules, forming a 'Baidu business ecosystem' in the era of large models.

This strategy mirrors the 'iOS' ecosystem. Apple stands at the pinnacle of global technology companies with this approach, generating revenue at an impressive rate. It would be disingenuous to suggest that Robin Li is not envious. He too aims to create his own 'iOS ecosystem' in the era of large models.

However, DeepSeek, like a catalyst, disrupted the global AI large model landscape and threw Robin Li's plans into disarray. As an advocate of the open-source route, DeepSeek's success is seen as a victory for open-source models. Some young AI scientists predict that there will be an increasing number of open-source large models in the future.

Many industry leaders have also started emphasizing the value of open sourcing. For instance, Alibaba co-founder and Chairman of the Board Joseph Tsai recently stated, 'DeepSeek has demonstrated the value of open sourcing.' If the goal of AI development is to create a closed-source system, to become the smartest PhD student, and win a Nobel Prize, I personally believe that such an approach holds little value.'

Their statements have amplified the merits of the open-source large model route promoted by DeepSeek. Consequently, Baidu has compromised, 'slapping in the face' by aligning with industry trends and shifting from closed sourcing to open sourcing.

'Leader's Dilemma'

In Tang Chen's view, Robin Li's change of heart, besides being influenced by DeepSeek, stems from a fundamental need: survival and an attempt to overcome the 'innovator's dilemma'.

As the first Chinese company to propose and enter the field of AI, Baidu has been researching and deploying in various AI sub-fields since 2013, with investments exceeding RMB 170 billion to date. Robin Li identified AI as Baidu's company-level strategy as early as 2016.

In March 2023, just three months after ChatGPT's launch, Baidu took the lead in introducing the 'Wenxin Yiyan' large model, staying ahead of the curve. Over the past two years, leveraging its first-mover advantage, Baidu has achieved notable success. According to Baidu's public data, as of November 2024, Wenxin Yiyan had a user base of 430 million, and the daily call volume of the Wenxin Large Model exceeded 1.5 billion, a more than 30-fold increase from 2023.

Furthermore, a report by the renowned research institution Frost & Sullivan recognized Baidu's position in the global AI industry's first tier, placing it in the same quadrant as Google and OpenAI as an AI-Native Giant.

However, Baidu's technological leadership has not translated into significant improvements in business and revenue. In the past two quarters, Baidu's revenue has declined noticeably. When Robin Li's advocacy of 'not competing in models, but in applications' became an industry consensus, Baidu's performance in 'AI super applications' under the closed-source model was underwhelming.

Data shows that Baidu is being overtaken by later entrants like Doubao and Dark Side of the Moon in C-end applications. According to January data from the AI Product Ranking, Wenxiaoyan's MAU (Monthly Active Users) was 13.05 million, ranking fourth after Doubao, DeepSeek, and Kimi.

This led to remarks by Shen Dou, Executive Vice President of Baidu Group and President of Baidu Intelligent Cloud Business Group, after the Baidu Cloud all-hands meeting on the 12th, sparking a 'dispute' with ByteDance's Volcano Engine business. Many commentators saw this as 'shifting blame' for subpar performance.

Baidu's predicament is precisely the 'innovator's dilemma' or the 'leader's dilemma.' The term was also used by 'VC Internet celebrity' Zhu Xiaohu when recently commenting on OpenAI.

The context is that OpenAI is also turning to open sourcing. Its founder, Sam Altman, recently acknowledged that the emergence of DeepSeek has altered the landscape where OpenAI had been far ahead for years. He even stated in a Q&A session: 'I think we are on the wrong side of history and need to develop a different open-source strategy.'

Zhu Xiaohu believes that Altman's admission of mistakes in the open-source strategy represents the 'leader's dilemma.' In the global AI landscape, DeepSeek has questioned the value of closed-source models. If closed-source models are costly and their performance advantages are not apparent, their significance diminishes, and OpenAI faces the same challenge.

Circumstances are stronger than individuals. Closed sourcing or open sourcing is merely a route choice. Under the impact of DeepSeek, Robin Li's swift strategic adjustment may not be a wise decision. However, it can also be seen as his attempt to overcome the 'leader's dilemma.'

A CTO of a listed company with deep cooperation with Baidu praised this move. He believes that Baidu's entry into the open-source model sphere is undoubtedly beneficial, as it can reduce the cost for enterprises to use large models and save usage expenses. Additionally, after Baidu open sources the Wenxin model, more developers will contribute to it, ultimately enhancing the model's capabilities and benefiting users.

Robin Li 'Acts According to the Situation'

Before officially open sourcing, Baidu has several months to adjust its relevant strategies. It can be predicted that Baidu will make even bigger moves in the future, including but not limited to changes in product routes, market strategies, and even organizational structure.

This time, Robin Li has embraced the essence of 'Lei Jun's School of Thought': acting according to the situation.

In 2007, after Kingsoft's IPO, Lei Jun was physically and mentally exhausted. He once said, 'The entire Kingsoft company was dragged down by the IPO, and this fatigue is difficult to express in words.'

This statement has dual meanings: one for Kingsoft, as it missed the internet's golden age while pursuing an IPO; the other for Lei Jun personally, as Kingsoft's IPO did not fully represent his personal success. Lei Jun, with top-tier qualities in all aspects, was successively outpaced by fellow entrepreneurs and became a 'supporting role' in the internet wave.

As described in Xiaomi's official biography, after leaving Kingsoft, Lei Jun pondered for over six months, 'We worked so hard, so why didn't we become mainstream in the internet wave?' The lesson Kingsoft taught Lei Jun can be summed up in four words: acting according to the situation. Lei Jun's personal signature on WeChat is also 'acting according to the situation.'

In fact, Baidu did not initially follow the closed-source route for large models. In 2021, Baidu open sourced the four pre-training models of Wenxin ERNIE, including the multi-granularity language knowledge enhancement model ERNIE-Gram, the long text understanding model ERNIE-Doc, the cross-modal understanding model ERNIE-ViL that integrates scene graph knowledge, and the language and vision integrated model ERNIE-UNIMO.

It was not until October 2023 that Baidu released Wenxin Large Model 4.0, and Wenxin Yiyan began charging users on November 1, 2023: Wenxin Yiyan Professional became China's first large model product to adopt a membership model for C-end charging, with a professional version priced at RMB 59.9 per month and a continuous monthly subscription discount price of RMB 49.9 per month.

Robin Li's stance is gradually softening. At an event in the United Arab Emirates on February 11, when directly asked about his views on DeepSeek, Li responded, 'Innovation cannot be planned. You don't know when and where innovation will come; what you can do is create an environment conducive to innovation.'

However, he also stated that he will continue to invest to ensure that Baidu remains at the forefront of this technological innovation or revolution and will utilize more computing power to explore different paths. Li believes that perhaps at some point, a shortcut will be discovered, such as training a model with only $6 million, but before that, billions of dollars may have already been spent determining which path is the right one to spend that $6 million.

Between the lines, it's evident: Robin Li has compromised due to reality but still adheres to his technological idealism. Baidu's foundation in the AI field may be his confidence in addressing external doubts. Acting according to the situation will also provide him with clearer goal setting, allowing him to escape the 'innovator's dilemma' and approach the debate over 'closed sourcing or open sourcing' with a more relaxed mindset.

Reference materials: China Entrepreneur Magazine, 'Robin Li is Also "Rolling Up"' by Tang Chen, 'Lei Jun is No Longer a Shadow, Finally Becoming "Lei Jun"'