Wenxin 4.5 ist ab sofort kostenlos! Ist die Goldgrube der nativen multimodalen Großen Sprachmodelle angebrochen?

When cost and capability are no longer in conflict, the implementation of AI applications is no longer an option but an inevitable event.

On March 16th, Baidu officially released the Wenxin Big Model X1 and Wenxin Big Model 4.5. Both models are now available on the official website of Wenxin Yiyan and are free for users.

Since the launch of the Wenxin Big Model 1.0 in March 2019, the iteration cycle of the Wenxin Big Model has been leading the domestic AI industry. After the release of the Wenxin Big Model 4.0 in October 2023, the user base of Wenxin Yiyan exceeded 100 million within two months.

Therefore, the already announced free Wenxin Big Model 4.5 and X1 are also expected to become the next fulcrum for domestic big models to break through the circle.

This time, both Baidu and the entire AI field aim not only to increase the number of users but also to realize the industrial value of AI applications.

01

On the Eve of the Application Explosion, Leading Model Manufacturers Continue to Compete in Basic Capabilities and Low Prices

From Grok to Gemini in Google AI Studio, the "protagonists" in the big model industry are constantly changing. The only constant is that all application fields are still on the verge of commercialization and marketization.

It is at this juncture that Baidu has successively released two cutting - edge models, indicating its next - step direction: to use the multi - modal barrier to embrace the likely AI application explosion in 2025.

As Baidu's next - generation base big model, the qualitative innovation of the Wenxin Big Model 4.5 lies in its native multi - modality. Through joint modeling of multiple modalities for collaborative optimization, its multi - modal understanding, text, and logical reasoning abilities have been significantly improved. It outperforms GPT4.5 in multiple tests, and the API call price is only 1% of that of GPT4.5.

The Wenxin Big Model X1 is a deep - thinking model benchmarked against the inference ability of DeepSeek R1. Its core goal is to upgrade the big model from a "generation tool" to a "decision - making brain". Currently, the price of X1 is only half of that of DeepSeek R1. Meanwhile, in the evaluation results, X1 shows comparable capabilities to DeepSeek R1 in many aspects.

For example, in the language generation evaluation, X1 can generate a script - killing - level suspense reasoning story based on a simple background description instruction.

The ability to imitate the "sharp comments" tone on Chinese social media is also a highlight in the evaluation. Compared with the "neutral" impression users have of big models, X1's answers are more opinionated, sharper in scenarios that require an attitude, and provide new ideas for the production efficiency of multi - channel content operations.

Since the release of Wenxin 4.0, in less than a year and a half, while investing in R & D, Baidu has also been committed to building an AI ecosystem: providing ready - to - use AI development support; deeply integrating the Wenxin Big Model into more than a dozen products such as search, maps, cloud storage, and document libraries. In 2021, Baidu entered a new stage of integrating cloud and intelligence from the transitional stage of "AI + cloud".

The low price or even free access to Wenxin 4.5 and Wenxin X1 is not just a technology and market strategy. Behind it is the cost reduction driven by technological innovation and a speed - up strategy for the application explosion stage. Baidu's action of releasing and freely opening Wenxin 4.5 shows three trends:

First, as one of the few global AI companies with a four - layer architecture layout, Baidu's R & D efficiency and its "foundation" at the chip layer give it full confidence in long - term marketization and commercialization.

Second, in addition to big models, domestic large enterprises are also competing in the MaaS platform. From now on, enterprises and developers can call the Wenxin 4.5 API on Qianfan, and Wenxin XI will be launched soon.

Third, in terms of the development stage of the entire domestic AI field, enterprises' enthusiasm for embracing new technologies is unprecedentedly high. As a leading manufacturer, Baidu intends to accelerate the ecological construction of AI industrialization by lowering the access threshold for developers and enterprises.

02

Who Defines the Productivity of Native Multi - modality?

From the following evaluation case of complex semantic understanding and online analysis, we can see that the significant improvement in the anti - hallucination ability is another outstanding feature of Wenxin 4.5.

The excellent performance of Wenxin 4.5 in anti - hallucination and other abilities also benefits from innovative technologies such as FlashMask Dynamic Attention Mask, multi - modal heterogeneous expert expansion technology, and big data construction technology with high knowledge density.

In our test, when we input a movie screenshot, Wenxin 4.5 can accurately identify its source. This kind of ability will be very useful in application scenarios such as intelligent search, content review, and film and television culture.

The ability to recognize and understand charts and then assist in decision - making can also reflect its multi - modal generation performance after anti - hallucination.

Actually, the industry's optimism about multi - modal big models has long been an open secret.

By starting from multi - modal joint modeling for images, videos, and texts, semantics can be uniformly expressed, thus solving the problem of information fragmentation in the chatbot mode of big models.

Recently, the intelligent design of vibe design based on the MCP protocol on blender has attracted a lot of attention. By adjusting the design model through multi - round conversations, the big model becomes a "controllable tool" that maintains consistency, rather than a student handing in homework.

Last week, the performance of multi - round conversations, image generation, and editing based on Gemini's capabilities on Google AI Studio also made it the "main AI" for many developers. Gemini has occupied the center of public opinion precisely because of its multi - modal generation performance.

Looking at the domestic market, Wenxin 4.5 is Baidu's first native multi - modal big model. The R & D achievements at this time are at the global forefront. In addition, these new trends indicate that the future trend of big models and AI industrialization is multi - modal productivity.

It is reported that Wenxin X1 also supports multi - modality and performs particularly well in Chinese knowledge Q&A, literary creation, manuscript writing, daily conversations, logical reasoning, complex calculations, and tool calls.

The capabilities of Wenxin X1 benefit from many powerful technologies. First, through progressive reinforcement learning, its application capabilities are comprehensively improved in scenarios such as creation, search, tool call, and reasoning. Second, for scenarios such as deep search and tool call, which most deep - thinking AI applications are not good at currently, Wenxin X1 has carried out end - to - end training based on the thinking chain and action chain and established a diversified and unified evaluation system.

In terms of cost, PaddlePaddle and Wenxin jointly achieved deep compression and inference acceleration in three aspects: model compression, inference engine, and system framework through quantization of the attention mechanism for long sequences, optimization of low - precision and high - performance operators, and a separated deployment architecture. This is also the technical logic source for Wenxin X1 to reduce the cost to half of that of DeepSeek R1.

03

Technology Inclusiveness to Promote AI Industrialization

As the global big models enter the "deep - thinking" era, the cost of big model industrialization has been continuously decreasing, which makes most enterprises no longer just onlookers of AI.

The core pain points for enterprises to promote the industrial application of AI currently focus on two aspects: high technical threshold and unaffordable cost.

Small and medium - sized enterprises are often limited by technical costs, while large and medium - sized enterprises, even if they have technical teams, often fall into the dilemma of "unbalanced input - output" due to high training costs and complex scenario adaptation.

For a long time before, the form of enterprise AI implementation and the understanding of technical boundaries were not clear, and there were indeed a lot of inefficient investments. However, with the improvement of the basic capabilities of major model manufacturers in the past two years, the generation effects of major models have made all industries start to embrace the possibility of big model industrialization.

Baidu's strategy of "low price + high openness" for Wenxin 4.5 and X1 directly addresses this contradiction.

In the past year, Baidu's intelligent cloud business has grown steadily and become the second growth engine, which also illustrates this point. More and more enterprises can use the capabilities of continuously iterated and updated basic big models to build applications through MaaS and tools integrated with basic big models without heavy upfront investment, and deconstruct the value of AI from the adaptability of scenarios and businesses.

The following is an audio - video content analysis evaluation of Wenxin 4.5. It can be seen that its performance is still remarkable in multi - modal processing based on factual answers without deep reasoning.

This case also confirms the feasibility of Baidu's idea of looking at the entire AI industrialization from the perspective of big model applications: the ready - to - use general capabilities and cost - effectiveness can turn AI industrialization from a heavy - asset investment into a lightweight tool. For example, users can extract audio and video from the cloud storage and summarize them in the Wenxin Big Model.

Undoubtedly, more powerful native multi - modal big models can release this potential in a more creative and lower - threshold way. With the features of free or low - price, tool call, and ready - to - use of Wenxin 4.5 becoming the standard for big models, the biggest obstacle to AI industrialization is no longer "whether it can be done" but "whether one dares to think".

In March 2023, Baidu was the first domestic big model manufacturer to propose "redoing all products with big models". Since then, Baidu has been firmly investing in training the next - generation basic models, including today's native multi - modal big models.

In 2025, all industries are optimistic that it will be an explosive year for the manifestation of AI application effects within enterprises. At the enterprise management and AI decision - making level, accuracy is actually very crucial in the large - scale use of big models. Therefore, Wenxin 4.5 has achieved a breakthrough in anti - hallucination. In terms of technologies such as RAG, which are closest to enterprise local deployment, Wenxin has been building search - enhancement technologies based on in - depth Chinese understanding in different industry scenarios such as the Internet, government affairs, and healthcare.

It can be said that from the perspective of Baidu's own business profile, whether it is data integration in vertical scenarios or RAG and iRAG, they are Baidu's advantageous fields accumulated over the years and also the core capabilities in big model competition.

For industrial implementation scenarios, comprehensive capabilities are often more important than having a single strong point. This is also one of the concepts of Agentic AI represented by Manus. Wenxin X1 is like a multi - faceted warrior that can expand its ability boundaries by autonomously calling tools, integrating callable tools such as search, document Q&A, and code interpreters.

In the field of AI applications, a positive trend is that big model manufacturers are evolving towards industrial implementation together with enterprises, rather than just engaging in a "competition" of parameters.

In 2025, Baidu has four moats: "search gene + technological barrier + manufacturer's engineering ability + first - mover advantage in the AI ecosystem". These past advantages and future development directions endow the Wenxin Big Model with the business gene of understanding and insight into local scenarios.

Where AI applications are headed, the evolution trajectory of technology seems to be clearly visible: from assistance to decision - making, from "labor force" to "brain", from single - point breakthrough to full - chain reconstruction. And Baidu's Wenxin, which has been "sharpening its sword for a year", has ushered in its big model year under the wave of technology inclusiveness.