Testing Baidu Search's DeepSeek full version: "Making it work for me" or "taking what's available"?
TechFlow Selected TechFlow Selected
Testing Baidu Search's DeepSeek full version: "Making it work for me" or "taking what's available"?
Baidu Search has now fully launched the full version of DeepSeek with internet connectivity, and over ten million users accessed it within just one hour of its release on the PC platform.
Navigating Web3 tides with focused insights
Image source: Generated by Wujie AI
Following Baidu Search's full integration of DeepSeek and the latest deep search capabilities from ERNIE on February 16, Baidu Search has now fully launched the full-capability version of DeepSeek.
Currently, users can enter any search query in the Baidu app, complete an initial search, then click the "AI+" button on the results page to access AI Search. After that, clicking the prompt below "Try the 'full-capability' version" enables direct conversation with DeepSeek. In addition, the PC version of Baidu AI Search has also integrated the full-capability version of DeepSeek R1 and provides web-connected functionality.
How effective is the full-capability version of DeepSeek in Baidu AI Search? After conducting dozens of tests, one cannot help but exclaim: Baidu AI Search is truly full-powered! With access to the best DeepSeek model capabilities, Baidu AI Search has opened the door to a new world.
"Utilizing for our own purpose," not "copy-paste"
After integrating DeepSeek, Baidu refrains from weakening it and ensures sufficient server resources. Behind this lies Baidu’s principle of "utilizing for our own purpose," rather than adopting a "copy-paste" approach. This process does not simply involve plugging DeepSeek directly into user-facing services—Baidu contributes its own "real substance."
What exactly is this "real substance"? It is directly reflected in Baidu’s deep technical integration and advantages in RAG technology, which help users solve practical issues such as large model hallucinations. At its core lies Baidu’s 25 years of data accumulation. As the world’s largest Chinese-language search engine, Baidu has amassed vast amounts of user search behavior data, webpage index data, and content ecosystem data. These datasets span multiple eras—from PC internet to mobile internet, and now the AI era—covering text, images, videos, and more, forming a unique "time barrier" that emerging AI search engines find difficult to replicate.
Let’s begin with a question that spans 25 years: Detail the development history of China’s new energy industry over the past 25 years, including specific time nodes, their impacts, and significance. The response from Baidu AI Search’s DeepSeek-R1 full-capability version is shown below:
The answer comprehensively outlines the 25-year development history of China’s new energy sector, dividing it into distinct phases, with some data points absent from other AI search engines.
During the search process, Baidu AI Search also demonstrates its new feature, the "Special Inspiration Zone," which provides inspirational content based on the current query, helping users organize their thoughts and gain a holistic understanding through a single question. Users can also click any suggested question within the inspiration zone to initiate the corresponding inquiry.
Notably, most current AI search products on the market display the model’s "thinking" process, typically shown as "searching X webpages." However, Baidu AI Search powered by the full-capability DeepSeek R1 can meticulously break down complex questions. For example, inputting the prompt: "Please analyze the changes in global GDP rankings between 2010 and 2015, providing specific values and illustrating the change process."
From verifying reliable data sources, defining whether the scope refers to nominal GDP or PPP-adjusted GDP, to retrieving historical data and performing rankings—the process mirrors how the human brain systematically dissects a problem.
Impressively, Baidu AI Search with the full-capability DeepSeek R1 thoroughly considers minute details and handles special cases thoughtfully—such as accounting for actual GDP growth rates across countries, fluctuations, exchange rate impacts, and rounding in data processing—something I did not expect. One can only exclaim: Baidu AI Search with DeepSeek R1 full-capability is indeed incredibly comprehensive!
The complete response to this question automatically adopts a table format without prompting, and clearly highlights key events.
After combining Baidu’s RAG with DeepSeek-R1, has hallucination truly been reduced? The next test examines the hallucination level of Baidu AI Search’s DeepSeek full-capability version: "Provide the global stock market closing data for February 20, 2005."
The trickiness of this question lies in the fact that February 20, 2005, was a Sunday, when most countries do not operate stock markets. A hallucinating model would fall into this trap, misattributing data from other dates to February 20, 2005. However, the thinking process of Baidu AI Search’s DeepSeek full-capability version is as follows:
As seen from the reasoning process, the system correctly identifies the issue during analysis, avoiding "plausible-sounding nonsense," and delivers a rigorous and accurate answer.
Building on this, let’s increase the difficulty: "Provide hourly data for European markets on February 18, 2005." This requires finer granularity in the response. The output from Baidu AI Search’s DeepSeek R1 full-capability version is shown below:
Baidu’s long-accumulated professional data is further unleashed through this integration. Asking a specialized question: "Please provide a detailed analysis of the mechanism behind China’s drug pricing and its influencing factors. In the context of current policy trends, explain the role and significance of medical insurance drug price negotiations."
This question targets pharmaceutical pricing in the healthcare industry, testing the model’s domain-specific knowledge depth and maturity. The explicit requirement to incorporate up-to-date policy context also evaluates the model’s ability to capture recent information. The resulting answer is shown below:
The answer includes numerous detailed data points—for instance, noting that annual treatment costs for PD-1 inhibitors dropped from 300,000 to 50,000–100,000 yuan under the "volume-for-price" scheme, and that CAR-T therapy prices fell from 1.2 million to 330,000 yuan among the 74 newly added drugs covered by medical insurance—indicating the model’s deep accumulation of data in the pharmaceutical field.
DeepSeek-R1 significantly enhances Baidu AI Search’s reasoning capabilities, while Baidu AI Search itself leverages its strengths by integrating multimodal abilities atop the DeepSeek large language model. Inputting the prompt: "I want to submit a painting to a graffiti illustration competition with the theme 'future city,' creating a fantastical world featuring elements like mysterious forests, magic castles, alien creatures, elves, and giants—please generate an image." This request involves multiple visual themes, requiring both linguistic understanding from the large language model and image generation from the visual model. The final generated image is shown below:
With the activation of the DeepSeek-R1 full-capability version, Baidu AI Search’s coding capabilities have also reached a new level. Inputting the prompt: "Write code that generates an animation celebrating Ne Zha’s box office success, including Ne Zha’s character, and verify the code’s effectiveness." The resulting output is shown below:
The response not only includes a complete executable code file but also provides verification instructions covering visual element checks, animation testing, and compatibility assessments, along with enhanced-effect code suggestions allowing users to customize adjustments.
Returning to user value, not technological silos
To date, looking back at the history of artificial intelligence development in China—whether it was Baidu’s early investment around 2010 in key technologies such as NLP, knowledge graphs, and machine learning, or its "first-mover" advantage during the large language model wave with the development of ERNIE Bot—Baidu’s consistent stance toward AI has always been "insisting on high-level self-research and increasing investment."
Even a company like Baidu, committed to self-reliance, has begun integrating external large models into its super app in response to the rise of DeepSeek, incorporating third-party model partners into its ecosystem. This marks progress toward an industrial operating system under its large model strategy—an ecosystem-level strategic combination centered on "large models + search."
This means Baidu Search is returning to its core focus on user value, refusing to become stagnant, and instead actively exploring integration between its core services and DeepSeek. Underlying this shift is a broader trend in China’s internet landscape: moving away from self-contained "siloed ecosystems" centered around super apps, toward an increasingly interconnected "technology community."
Join TechFlow official community to stay tuned
Telegram:https://t.me/TechFlowDaily
X (Twitter):https://x.com/TechFlowPost
X (Twitter) EN:https://x.com/BlockFlow_News
