On February 13, Baidu announced that its ERNIE Bot will be completely free to use starting April 1.

The following day, Baidu further announced that the next-generation ERNIE Bot 4.5 series will be open-sourced from June 30.

Affected by these announcements, Baidu Group's share price surged 5.74% on February 13, closing at HK$94 per share. Notably, Baidu was an early pioneer in large language models in China. Previously, ERNIE Bot offered a free basic version and a paid professional version priced at RMB 59.9 per month, with a discounted rate of RMB 49.9 for continuous monthly subscriptions.

Today, domestic large models like Tongyi Qianwen, Kimi, Doubao, Tencent Hunyuan, and DeepSeek are proliferating rapidly. Has ERNIE Bot, with its early start, been able to maintain its lead?

01

Once a Pioneer in the Large Model Race

ERNIE Bot: As the earliest AI language model in China, some say that for you, all others are latecomers. Why do you think that is?

Following the explosion of ChatGPT's popularity, the significant boost to personal productivity and the enormous commercial potential of large models across various industries have become increasingly evident. In February 2023, just as Baidu publicly launched "ERNIE Bot," partners across its various products, including Baidu Search, Intelligent Cloud, and Autonomous Driving, immediately expressed their need for such a product. Shortly thereafter, over 600 partners announced their integration into the ERNIE Bot ecosystem, eager to leverage the latest and most advanced large language model.

On March 16, 2023, Baidu's new-generation large language model, ERNIE Bot, officially began invitation testing. By August 31, ERNIE Bot had taken the lead in becoming fully publicly accessible. On the first day of its openness, ERNIE Bot responded to over 33.42 million questions from netizens. On December 28, Baidu Chief Technology Officer Wang Haifeng announced at the 10th WAVE SUMMIT Deep Learning Developer Conference that the user base of ERNIE Bot had surpassed 100 million.

Li Yanhong commented at the time: "ERNIE Bot will change the rules of the cloud computing game."

When Baidu executives candidly admitted that ERNIE Bot was not yet fully mature, Baidu's share price dropped accordingly. Li Yanhong openly noted that his experience with the model was not perfect and that such large language models were still far from fully developed. Wang Haifeng also mentioned that the model was currently "insufficiently trained."

Fast forward to 2025, it appears that ERNIE Bot has yet to overcome these "imperfections." When I posed this question to it, ERNIE Bot provided a detailed response, summarizing that "although as an early AI language model, I have a certain historical position, in the context of continuous technological and application development, later language models have indeed gained significant advantages in many aspects."

02

Early but Late?

Despite its early start, is "ERNIE Bot" truly ahead?

Baidu has been deeply invested in the field of AI for over a decade, developing a globally rare IT four-layer technology stack architecture comprising "chip-framework-model-application." This spans from high-end chips like Kunlun, to the PaddlePaddle deep learning framework, to the Wenxin pre-trained large model, and finally to applications such as search, intelligent cloud, autonomous driving, and DuerOS. Each level boasts leading self-developed technologies and products.

In terms of advantages, ERNIE Bot leverages Baidu's substantial data and technological accumulation, boasting a vast Chinese corpus and certain advantages in Chinese understanding. As mentioned earlier, it has over 600 partners, and Baidu actively deploys application ecosystems, collaborating with multiple enterprises to facilitate product implementations.

Computing power chips are crucial. In June 2021, Baidu's intelligent chip and architecture department completed independent financing, establishing Kunlun Core (Beijing) Technology Co., Ltd., with a first-round valuation of approximately RMB 13 billion. Ouyang Jian, Baidu's Chief Chip Architect, assumed the role of CEO of Kunlun Core. A year earlier, Baidu had already trial-produced the first-generation cloud-based general-purpose AI computing processor "Kunlun 1" chip, which was the only cloud-based AI chip in China at the time to undergo large-scale core algorithm testing on the internet. Two months after the independent financing of the chip department, Baidu announced the official mass production of the second-generation Kunlun chip, "Kunlun 2."

It is evident that Baidu's AI layout is ambitious and grand. "ERNIE Bot" is merely a small part of the research and development process based on Baidu's Wenxin large model. This can be glimpsed from the content presented on the Baidu Wenxin official website. In the overall planning of the Wenxin large model, the dialogue function occupies only a minor portion. Imagine if ChatGPT had not suddenly gained global attention, perhaps "ERNIE Bot" would have remained relatively quiet on the project schedule for a considerable time, with limited chances of coming into the public eye.

In contrast, DeepSeek's success does not seem coincidental.

In terms of the most intuitive download volume, DeepSeek topped the download rankings in 140 market app stores, achieving 16 million downloads in just 18 days, outperforming ChatGPT's performance over the same period.

The DeepSeek - R1 model training utilized approximately 2,000 NVIDIA dedicated chips, fully leveraging the powerful parallel computing capabilities and CUDA ecosystem of NVIDIA chips to achieve efficient deep learning calculations. In terms of hardware adaptation, DeepSeek is also actively exploring integration with domestic hardware, such as Ascend and Haiguang chips. Technologies like sparse computing enable it to perform well on domestic hardware, reducing dependence on specific foreign chips.

Due to its innovative architecture design and training strategies, DeepSeek's chip demand for tasks like inference is relatively flexible, requiring only a hundred-card scale during training.

Besides DeepSeek, platforms like Keling, Jimeng, and Doubao may adopt lightweight models or customized architectures, focusing on optimizing interactive experiences and specific scenarios (such as creative generation and dialogue interaction). Overall, DeepSeek excels in complex tasks and technical scenarios, while ERNIE Bot shines in Chinese processing and cultural understanding. Keling, Jimeng, and Doubao are more suitable for creative generation and lightweight interaction.

03

Can Baidu Catch Up to the Late Crowd?

Prior to its open-source announcement, netizens criticized ERNIE Bot for pursuing short-term commercialization while neglecting in-depth technical research and development.

ERNIE Bot's deficiencies are indeed apparent. Compared to leading models at the technical level, there are gaps in logical reasoning, creative text generation, etc. At the application level, practical scenarios are limited, the user experience needs improvement, and there is a lack of impactful application cases. At the ecological level, the construction of the open-source ecosystem lags, and developer community activity is low.

Baidu's ERNIE Bot announced that it will lift the fee restriction from April 1, 00:00, making its latest model and all advanced features accessible to all users. This decision marks the official entry of domestic large model services into the stage of universal benefit.

ERNIE Bot simultaneously launched a deep search function. This new feature boasts enhanced thinking, planning, and tool invocation capabilities, enabling it to provide users with expert-level content responses and handle multi-scenario tasks, realizing multi-modal input and output. Currently, users can be the first to experience this function on the ERNIE Bot official website, with related mobile App functions to be launched soon.

This move not only reflects the technological iteration and cost optimization breakthroughs made by the Wenxin large model but also demonstrates Baidu's strategic layout and market competition determination in the AI field. With the implementation of the free and open policy, the application threshold for AI technology will be further lowered, significantly promoting the popularization and innovative development of artificial intelligence technology.

Li Yanhong was originally a staunch supporter of closed-source large models. On occasions like the 2024 World Artificial Intelligence Conference, he repeatedly expressed his support for closed-source models.

From a performance and cost perspective, Li Yanhong believes that under the same parameter scale, open-source models are inferior to closed-source models in capability. If open-source models aim to catch up with closed-source models, they require a larger parameter scale, which translates to higher inference costs and slower response speeds. Closed-source models are typically optimized by professional teams for specific purposes and are tightly integrated with specialized hardware, enabling economies of scale and lower inference costs.

While acknowledging that open-source large models have value in specific scenarios such as academic research and teaching, Li Yanhong emphasized that in a fiercely competitive commercial environment, businesses must strive for higher efficiency and lower costs than their peers. Here, commercialized closed-source models excel.

Li Yanhong once stated that only closed-source models can sustain a true business model, attracting talent and computing power. Outstanding talents are more inclined to choose teams with closed-source models because they offer better resources and development opportunities.

The rise of DeepSeek and others has put pressure on ERNIE Bot, and open sourcing can attract more developers and users, enhancing Baidu's market competitiveness. Furthermore, with technological advancements, the inference cost of large models has dropped significantly, giving Baidu the means to open source and promote wider technology application.

Is AI headed towards a price war? ERNIE Bot announced free access simultaneously with OpenAI. At 3:00 am Beijing time on February 13, OpenAI CEO Sam Altman also announced that GPT-4.5/5 will be released soon, and the free version of ChatGPT will be able to use GPT-5 for unlimited dialogue under standard intelligent settings. It is reported that GPT-4.5, soon to be released by OpenAI, is the legendary Orion model and the last non-chain-of-thought model. Subsequently, it will integrate the two major models of GPT and O series to create a new system that can automatically select thinking and non-thinking functions, suitable for multiple tasks. Altman stated that this will be OpenAI's last "non-chain-of-thought model." Unlike other reasoning models from OpenAI, non-chain-of-thought models are often less reliable in fields like mathematics and physics.

DeepSeek, which promotes two closed-source models, offers this perspective on ERNIE Bot's comeback:

Baidu's stock prices surged in response, which is good news for large-scale deployment among enterprise-level users. However, whether it can become the super app envisioned by Li Yanhong remains to be seen.