Wenxin Big Model Open Source: What are the important signals and impacts released?

Introduction:
・ERNIEOpen Source Background:AI modeling trends, technology accumulation and breakthroughs;
・The significance of open source and openness:Promote technological advances, lower thresholds, and ecological construction;
・Open source and open content:Models and frameworks, open source protocols;
・The Impact of Open Source and Openness:Industry applications, international competitiveness;
・The Open Source Openness Challenge:Intellectual property protection, data security and privacy;
・Future Outlook:Technological innovation and application expansion.

On Feb. 14, the country's headAI macromodelcompanyBaidu("Baidu") announced that in the coming months it will be launching a series ofthe train of thought of a writerLarge Model4.5系列，并于6月30日起正式开源.

According to Baidu's official introduction, the Wenxin Big Model 4.5 series has been comprehensively upgraded in many aspects, including algorithm optimization, model architecture, and data processing. Compared with its predecessor, the new version has a significant improvement in performance, which can better handle complex tasks and provide more accurate and intelligent services. In addition, Wencent Big Model 4.5 series also introduces more innovative technologies, aiming to meet the increasingly diverse needs of users.

Just the day before, Baidu also announced that Wenxin Yiyin will be fully free of charge from 0:00 on April 1, and all PC and APP users can experience the latest model of the Wenxin series, which is further proof that Baidu is gradually abandoning the once commercialized closed-source strategy.

It is worth noting that also in the past two days, from OpenAI to Google, also increased the openness of the big model, announced the news of free openness of its big model products.

OpenAI is also expected to go open source. It is reported that OpenAI is currently discussing matters such as publicizing the weights of AI models internally.

The AI giants are moving in lockstep towards open source and openness, releasing a strong signal:

Two years after the big model rampage, the big model technology in the B, C two-end landing paradigm, emerging a new change, the big model vendors put forward higher requirements - they not only want to walk in the forefront of the big model technology, but also in the big model application outbreak of the eve, to accelerate the exploration of the big model landing of the cost reduction path, the first to take the lead in the run.

The open source and openness of the Baidu Wenshin Big Model is based on the above two points.

On the one hand, in the big model wave of the past two years, Baidu is one of the AI companies that have invested the most, have the fastest technology iteration, and have the widest and deepest exploration of B-end industry landing and C-end application. As of November 2024, the user scale of Wenxin Yiyin was 430 million, and the average daily call volume of Wenxin's big model was more than 1.5 billion times, which was more than 30 times higher than in 2023.

On the other hand, from model inference to model training, Baidu has realized an effective reduction in cost through technological innovation. And when Baidu, OpenAI, Google and other AI giants take the lead to turn, when a greater degree of technology, ecological openness has become an industrial consensus, AI technology inclusive, is accelerating into reality.

From To C to TO B, why open source and openness have become the way to go for the big model industry?

For the past two months or so.DeepSeek's stones thrown at the big model industry continue to ripple.DeepSeek-v3The low cost on large model training presented,DeepSeek-R1The low cost in model reasoning and the stunning ability in thinking logic, Chinese language, programming, etc. presented by the DeepSeek app quickly fueled it to become the most globally recognized AI company around the Chinese New Year.

And its demonstration of the possibility of reproducing advanced models at low computational costs, the explosion of DeepSeek applications, somehow confirms one thing:

The big modeling industry, at present, has entered a new phase that requires open source and openness.

From an objective point of view, open source and closed source, these two different technological routes are not completely opposed to each other, only that they will present different characteristics at different times of industrial development.

For example, the early model open source is more like marketing, Meta's Llama chose to semi-open source, only open source part of the parameters and configuration files, but this to a certain extent will instead affect the verifiability and credibility of the model.

However, by this year, the big model is accelerating into the AI application explosion stage after passing through the initial development stage, a stage where the open-source route is clearly more conducive to the spread of the big model technology and increasing the adoption rate.

As Robin Li said, "At the end of the day, the most important thing is the application, not which big model is used. Whether it's open source or closed source, what's more important is what kind of value can be created at the application layer."

For example, in the B-side market, the "2024 China Enterprise AI Big Model Application Status Research Report" points out that the penetration of AI big models in the enterprise is still in the early stages, although 55% enterprises and organizations that have already deployed big models believe that they have already seen the clear business value brought by big models.

The problem is that for many enterprises, especially small and medium-sized enterprises (SMEs), cost, technology, talent and industry solutions, remain several major challenges in the process of landing big models, and they remain positively and cautiously ambivalent about investing in AI big models.

IDC also mentioned in the "China SMB Generative AI and Large Model Adoption Survey" that the cost of hardware, software, training and data processing required to adopt large models and AI technologies is also a challenge for many SMBs.

Focusing on the C-end market, although the industry has not yet appeared a real super application, but the user's habit of using large model applications is accelerating the formation of a comprehensive openness, but also a general trend.

That is to say, full open source and openness, in order to better meet the B-end enterprise customers, C-end users constantly growing market demand.

We have seen that when the wind shifted, big model head players such as Baidu and OpenAI, keenly captured the signals and took the lead to open source and open up in a more aggressive manner.

Baidu, for example, in addition to the full openness of Wenxin Yiyin in the C-end, in the B-end, Baidu in the big model ecology is also gradually increasing openness.

On February 3, Baidu Intelligent Cloud officially announced that the DeepSeek-R1 and DeepSeek-V3 models have been shelved on its Chifan ModelBuilder platform.

It is worth noting that Baidu has knocked down the price of these two models -- the price for customers to call these two models on the Chifan ModelBuilder platform is only 30% off of the official publication price of DeepSeek-V3 and 50% off of the official publication price of DeepSeek-R1, and at the same time provides a limited-time free The service is also free for a limited period of time.

On the other hand, over the past year, Wenxin's flagship large model has seen price reductions of more than 90%, and the main model is also completely free of charge, minimizing the cost of trial-and-error innovation for companies.

Of course, more importantly, for the next upcoming launch of the latest Wenxin Big Model 4.5 series, Baidu will also be officially open source from June 30th - it will face the market with a more positive attitude, and work together to promote the development of the industry.

Hearing Tide TI also noted that, from the information released so far, Baidu's open stance, to be more active than OpenAI - OpenAI consider open source, is previously released AI models, while Baidu's open source action is focused on the next upcoming release of the latest series of models.

This means that for the next wave of large model AI application outbreaks, Baidu is already running ahead of schedule.

Based on technological innovation, Baidu runs through the cost reduction path of big model technology

"Looking back over the last few hundred years, most innovations have been related to cost reduction, not just in AI, not even just in IT." So said Robin Li at the World Governments Summi2025 Summit in Dubai on February 11th.

In his view, if you can reduce costs by a certain amount, a certain percentage, that means productivity will increase by the same percentage, "and I think that's pretty much the essence of innovation. And today, the pace of innovation is much faster than it used to be."

Baidu founder Robin Li, Photo/Baidu official microblogging 

Behind Robin Li's statement, today's Baidu, has run through the big model technology cost reduction path. And the support behind it is technological innovation.

Specifically, from large model training to reasoning, Baidu's current cost reduction effect is more significant.

First look at the cost of training. Baidu's self-developed Kunlun core chip and the completion of the Wanka cluster provide arithmetic support for large model training, and the Baige-AI heterogeneous computing platform, which can carry the processing of large amounts of data, the training of oversized models, and the reasoning of high concurrency services, accelerates the AI tasks, and is a more basic layer of infrastructure.

One of the performance advantages of the Kunlun Core is that it can run large-scale models with fewer computational resources, which in turn allows for a reduction in the amount of computation required for inference and training of large models, directly reducing the arithmetic cost;

The advantage of large-scale clusters lies in the fact that they can improve the utilization of computing resources through task parallel scheduling, elastic arithmetic management, etc., avoiding idle arithmetic, improving the computational efficiency of a single task, and reducing the overall arithmetic cost. Recently, Baidu Intelligent Cloud successfully lit the Kunlun Core III 10,000-card cluster, which is the first officially lit self-developed 10,000-card cluster in China, and Baidu plans to further expand to 30,000 cards.

In addition, Baidu has also realized the efficient deployment and management of large-scale clusters, supported by the capabilities of the Baige platform.

For example, it has increased the effectiveness of bandwidth to more than 90%, effectively reduced the energy consumption of model training through innovative heat dissipation solutions, and increased the cluster MFU (GPU resource utilization) for training mainstream open-source models to 58% by continuously optimizing and refining the distributed training strategy for models.

Then look at the reasoning cost of the model. Some industry insiders have analyzed that one of the biggest reasons behind this full openness of Wenshin Yiyi is perhaps the decreasing cost of reasoning.

"Baidu has a comparative advantage in model inference deployment, especially with the support of the Flying Paddle deep learning framework, in which parallel inference and quantitative inference are Flying Paddle's self-developed technologies for large model inference. The joint optimization of Flying Paddle and Wenxin can achieve improved inference performance and reduced inference cost." He further analyzed.

Specifically, Baidu is the only AI company in China that has a four-layer AI full-stack architecture of "chip-framework-model-application", which means that Baidu has the thickest and most flexible technological base in China, and is able to realize end-to-end optimization, which not only dramatically improves the efficiency of model training and inference, but also further reduces the overall cost. This means that Baidu has the thickest and most flexible technology base in China, and is able to realize end-to-end optimization, which not only dramatically improves the efficiency of model training and reasoning, but also further reduces the comprehensive cost.

As an example, the lower inference prices of DeepSeek-R1 and DeepSeek-V3 on the Chifan ModelBuilder platform are precisely based on technological innovation -- Baidu Intelligent Cloud's inference engine performance optimization technology, inference service engineering architecture innovation, and inference service all-link security guarantee The deep integration of these technologies is an important reason for bringing down the price.

Based on the above points, Baidu's cost reduction path is actually particularly clear - based on self-research technological innovation, to improve the resource utilization of large models in the training and reasoning process.

We're also seeing that following the cost reduction path of this big model technology, Radish Express, is also accelerating to land at a lower cost.

In May last year, Radish Express released the world's first large model to support L4-level automatic driving, further improving the safety and generalization of automatic driving technology, using the power of the large model to make automatic driving "faster on the road", and the ability to deal with complex traffic scenarios, completely no less than Waymo.

Focusing on the sixth-generation unmanned vehicle of Radish Express, it comprehensively applies the program of "Baidu Apollo ADFM model + hardware products + security architecture", and ensures the stability and reliability of the vehicle through the 10-weight safety redundancy program and the 6-weight MRC safety strategy, and the safety level is even close to that of the domestic large aircraft C919.

Notably, this process has resulted in the Radish Express unmanned vehicle's cost, which is at or near the lowest level in the industry. Its sixth-generation unmanned car, than Tes plans to mass-produce in 2026 cybercab cost is lower, or even 1/7 of Waymo.

This somehow also accelerated the process of landing the Turnip Express.

Up to now, turnip fast running has been in the north, Guangzhou and Shenzhen and other cities, as well as Hong Kong, China to open the road test. Baidu previously revealed that the turnip fast running cumulative orders have exceeded 8 million orders. Robin Li also mentioned that the turnip fast running L4 level autonomous driving safety test mileage has accumulated more than 130 million kilometers, the risk rate is only 1/14/ of human drivers.

At the same time, the test mileage accumulated by Radish Express in the more complex urban road conditions in the Chinese market has also laid the groundwork for it to explore emerging markets such as the Middle East and Southeast Asia.

What's next for Baidu in the year of the app explosion?

"We live in very exciting times. In the past, when we talked about Moore's Law we said that every 18 months performance would double and costs would halve; but today, when we talk about big language models, we can say that the cost of reasoning can be reduced by more than 90% every 12 months. That's much faster than the computer revolution we've experienced in the last few decades." So said Robin Li at that Feb. 11 summit.

In fact, looking back at the dynamics of the big model circuit over the past year, from the price wars to the diverging paths of the big model vendors to the kimi's out of the ring to theAI AgentThe explosive momentum of DeepSeek, to the meteoric rise of DeepSeek and the resultingLarge model open sourceOpen Tide, it's not hard to find:

At present, the big model industry is accelerating into a new cycle - the speed of technology iteration is getting faster and faster, the unknown imagination of technological innovation is broader, the speed of cost reduction of big model technology is faster, and the breaking point of big model application is closer.

This also means that, from the perspective of market competition, the next big model vendors to compete in the dimension, will also be richer.

They have to fight for technological innovation, ecological empowerment, as well as openness and cost reduction capabilities, and applications.

However, with reference to Baidu's cost reduction path, in the long run, the most core competition, still focusing on one point - who can continue to walk in the forefront of large model technology innovation.

We note that this is also Baidu's long-term thinking.

"Innovation cannot be planned. You don't know when and where innovation will come, all you can do is to create an environment that is conducive to innovation." So says Robin Li.

This corresponds to the fact that, despite technological advances and technological innovation in the continuous cost reduction, Baidu will next continue to invest heavily in chips, data centers, and cloud infrastructure to create a better and smarter next-generation and next-generation models.

Baidu, for example, is still enriching its big model matrix.

Currently, Wenxin's large model matrix includes flagship large models such as Ernie 4.0 Turbo, lightweight models such as Ernie Speed, and also a series of thinking models and scenario models produced based on the base model to meet the needs of different applications.

In the third quarter of last year, Baidu also launched two enhanced lightweight models, Ernie Speed Pro and Ernie Lite Pro. Then this year, from the news that has been released, Wenxin big model 4.5 series and 5.0 series will also be released.

On the other hand, we also see that behind Baidu's more active open-source openness posture, in fact, continues to continue the previous concept - to accelerate the process of promoting the application of large models in B-end business scenarios, as well as the exploration of C-end applications.

Finally, as Robin Li says, "Maybe, at some point you'll find a shortcut to train a model for, say, only $6 million, but before that, you've probably spent billions of dollars on discovering which is the right way to spend that $6 million."

For Baidu, continuous high-pressure strong type of technology investment to create an innovative environment, in fact, is a "stupid work", but the good thing is that this is robust enough, solid enough, the potential opportunities are also greater.

For one thing, the call volume of Wenshin's big model was already the highest in China before, and now after open source, its call volume is expected to increase significantly, further expanding the use of Wenshin's big model;

Secondly, from the point of view of the big model ecology, Baidu has already established an ecological advantage based on an open stance in the past.

For example, Baidu launched the open-source Flying Paddle framework as early as 2016; Baidu's Thousand Sails big model platform, which also has the largest number of access models in the industry at present, supports nearly 100 mainstream models at home and abroad.

It can be predicted that now, after greater efforts to promote the big model open source, open, in the new round of big model competition, Baidu has begun to steal the show.

Article content partially sourced from commercialNew Knowledge No.: Hear the Tide TI reports