How did China's AI achieve a counter - attack against the US in just three years?

Artificial Analysis is an authoritative institution specializing in AI competency assessment.

It helps engineers and enterprises accurately gauge AI capabilities to make strategic decisions. Its approach is to rely on a large amount of data and conduct systematic analysis.

In mid - June 2025, they released a report titled: "Highlights Report of Chinese AI in Q2 2025".

This report assesses China's position in the global AI landscape and also compares the strengths of China and the United States in the field of language models.

01

Let's first look at some specific data and cases. Since the release of ChatGPT in 2022, the gap between Chinese and American AI seemed significant at one point. However, by May 2025, this gap had shrunk to less than three months.

This indicates that Chinese AI labs and enterprises have made remarkable progress in recent years.

DeepSeek R1 (May 2025) scored 68 points in the Artificial Analysis Intelligence Index; Alibaba's Qwen3 235B A22B scored 47 points.

In comparison, OpenAI's o3 also scored 68 points, and Gemini 2.5 Pro scored 67 points. As you can see, the gap between China and the United States in this regard has become very small.

How did China achieve this leap? To put it simply: improvement in technical performance and strategic breakthroughs.

Take DeepSeek's R1 model as an example. Through reinforcement learning (RL) optimization, it has significantly improved its reasoning ability. When this model was released in May 2025, it was already one of the most intelligent open - source models globally.

It not only has a parameter scale of 671B, with 37B active parameters, but also performs excellently in multiple evaluation criteria, such as MMLU - Pro and GPQA Diamond.

Now, let's look at the strategic breakthroughs.

Leading Chinese labs generally adopt an open - weight strategy, which means making the weights of their flagship models public to accelerate technology sharing and dissemination. This approach stands in sharp contrast to the closed - source strategy of leading American companies.

For example:

Alibaba's QwQ 32B Preview surpassed Meta's Llama 3.1 405B for the first time in November 2024, becoming the most intelligent open - source model at that time.

This open attitude not only promotes technological progress but also attracts the attention and participation of global developers. Moreover, Chinese enterprises are constantly exploring new application scenarios and technological paths.

For instance, Alibaba is not limited to language model research and development. It has also launched models covering multiple modalities such as image generation and video generation. This multi - modal layout enables the enterprise to maintain competitiveness in different fields.

Therefore, the narrowing of the gap between the cutting - edge language models of China and the United States is primarily due to technological innovation. However, in addition to technological breakthroughs, what truly enables Chinese AI to advance further and faster is actually a "strategic trump card": open weights.

What are open weights?

Some companies keep their most powerful models under wraps, fearing that others will find out. In contrast, some companies not only make their models public but also expose their "brains" for everyone to see.

Doesn't it sound a bit incredible? After all, models are trained at a high cost and over a long period. Aren't they afraid that others will use them if they are directly made public?

However, many Chinese AI labs and enterprises have chosen this path. Why? Because they have discovered that working in isolation is not as effective as opening the door to welcome others.

02

This "open" approach has three advantages. Firstly, it significantly lowers the technological threshold.

Do you want to engage in AI development? No problem. Simply download a pre - trained model and optimize, adjust, or even innovate on it. It's like building blocks; others have already laid the foundation, and you just need to add bricks on top.

Secondly, it accelerates technological dissemination.

Once a model is open - sourced, developers around the world can participate, offer suggestions, modify code, and develop new applications. As a result, the iteration speed of the model itself will become extremely fast.

The most crucial point is that it builds an ecosystem. Those who can make good use of it can create value on this platform, which in turn promotes the development of the model itself.

The Chinese AI community has truly grasped the concept of "open weights".

ByteDance's Seedream 3.0 (December 2024), an open - source model in the field of image generation, has not only been widely adopted in China but also attracted a large number of overseas developers for secondary development.

Tencent's Hunyuan 2.0 (April 2025), in the field of natural language processing, has become the preferred model for customized services of domestic enterprises after its weights were fully made public.

In addition, leading institutions such as Alibaba and Huawei have successively launched their flagship open - source models in the past year, and they even made the weights public.

These actions are not accidental but represent a consensus in the Chinese AI community: openness is not about giving up control but a way to gain influence.

However, what exactly is the gap compared to the United States?

Currently, companies like OpenAI, Anthropic, and Google still adhere to a "closed - source" strategy. Although they may release some API interfaces or model inference capabilities, the core model weights are not made public.

There are, of course, commercial considerations behind this, such as protecting intellectual property, maintaining competitive advantages, and preventing model abuse.

From another perspective, it also means that their models can only be optimized by internal teams or authorized users, and it is difficult for external developers to truly participate.

This is like an experiment in "technological democratization"; the more developers a party has, the stronger its future influence will be.

Is there any risk in doing this? Of course.

Firstly, there is the issue of the business model. If a model is completely open - sourced, how can one make money? The traditional methods of "selling models" and "charging licensing fees" are obviously no longer applicable.

Therefore, domestic enterprises are now exploring new models, such as providing advanced technical support, customized services, and commercial plug - ins, to make up for the revenue gap caused by open - sourcing.

Secondly, there are security risks. Once a model is made public, it is difficult to control its usage boundaries. If someone uses it to generate false content, conduct malicious attacks, or even carry out automated fraud, it will be troublesome.

However, these problems are not unique to China; they are new challenges that the entire industry has to face. The key lies in how to establish reasonable regulatory mechanisms and ethical norms.

So, you can see that China's "open weights" strategy is more like a strategic mindset. By lowering the threshold, accelerating dissemination, and building an ecosystem, it has transformed AI technology into an infrastructure that more people can participate in and benefit from.

And this is also one of the important reasons why Chinese AI has been able to catch up rapidly in just a few years and even surpass the United States in some fields.

03

Technological innovation and open weights can help China catch up with the United States quickly. However, the question arises: how can technology be transformed into products? Who is driving these models to be truly implemented and used by the public?

The answer lies in the "enterprise ecosystem" of Chinese AI.

This is a group of companies that are leveraging their respective strengths to form a collaborative ecosystem. It's a bit like a sports team, where some are responsible for organizing attacks, some for defensive counter - attacks, and others are ready to come on as substitutes at any time. This is how China's AI industry has emerged.

Let's first talk about the first type of players: large technology companies, such as Alibaba, Tencent, Huawei, and Baidu.

What roles do they play in the ecosystem? To put it simply, they build platforms, provide computing power, and develop underlying models. You can think of them as the "infrastructure construction teams" in the AI era.

Alibaba not only trains large models on its own but also launched the ModelScope platform, which allows developers to easily download, test, and deploy various AI models.

Huawei launched the Pangu series of large models and provided supporting chips and cloud services to offer underlying support for the entire industry. After Tencent's Hunyuan series of models made their weights public in April 2025, they quickly became one of the preferred models for customized services of domestic enterprises.

These large companies are building an open, shared, and scalable technological platform to enable more people to participate.

Now, let's look at the second type of players: AI startups. The number of such companies is increasing, and they have one thing in common: they don't aim for comprehensiveness but for excellence in a specific area.

For example:

Moonshot AI focuses on long - text processing. Its Kimi model supports ultra - long contexts and is suitable for tasks such as document analysis and legal assistance. MiniMax specializes in lightweight models, aiming for efficient operation on mobile phones and edge devices, which is more suitable for small and medium - sized enterprises.

The existence of these startups makes the entire AI ecosystem more diverse. Instead of trying to do everything like large companies, they dig deep in one direction, making it easier for them to develop unique features and competitiveness.

The third type of players is the most easily overlooked but also the most down - to - earth: cross - border players.

What does "cross - border" mean? Companies that were not originally in the AI field are now starting to use AI as a core capability.

Xiaomi, which used to focus on mobile phones and smart home devices, has now launched its own AI model, MiMo - 7B, and integrated it into Xiaoai, its voice assistant, making the voice assistant smarter and more user - friendly.

Baidu, a company that started with search engines, is also continuously iterating its Wenxin Yiyan, trying to integrate AI into multiple business lines such as search, advertising, and autonomous driving.

360 has even more directly launched a "Brainy" product matrix, integrating AI into existing products such as security software, browsers, and office tools, making AI more accessible.

The advantage of these companies lies in their user base, application scenarios, and implementation experience. They don't need to start from scratch. Instead, they use AI as an "enhancer" and add it to their existing products and services to create real value.

So, you can see that these three types of players have different divisions of labor:

Large companies provide underlying models and platform support; startups focus on model innovation and technological breakthroughs; cross - border players are responsible for rapid implementation and commercialization.

In this way, it becomes a technological platform jointly built by multiple parties and accessible to everyone.

The development of Chinese AI relies on such a multi - level, multi - role, and collaborative enterprise ecosystem. It is this ecosystem that enables China to forge its own path when facing a powerful opponent like the United States with a strong technological foundation.

04

Where will this path ultimately lead? Will the AI competition between China and the United States lead to cooperation or complete confrontation?

For AI to truly enter our lives, it's not just about being able to chat and write articles; it also needs to understand images, recognize voices, and even generate videos. This is the "intelligent experience" we desire.

In other words, future AI must be "multi - modal", capable of understanding multiple types of information such as text, images, sounds, and movements, and freely switching between these modes.

So, can China's multi - modal capabilities compete with those of the United States?

In the field of language models, judging from historical data, the gap between China and the United States has become very small, and they can be said to be on the same level.

The field of image generation is an area where foreign companies started earlier. The names Midjourney and Stable Diffusion are almost well - known in the designer community.

However, in recent years, Chinese models have gradually caught up.

ByteDance's Seedream 3.0 became popular soon after its release at the end of 2024. Its image quality and detail restoration can rival those of Midjourney V6. Its ELO value reached 1111, only a little short of V6's 1150.

Alibaba's Seedream series is also continuously being updated, with special emphasis on supporting Chinese prompts, which is more in line with the usage habits of domestic users.

What does this mean? China and the United States are on the same starting line in this field, and the winner will be the one who can meet user needs more quickly and implement the technology in actual products.

The most challenging area is video generation.

If image generation is a "detail - oriented competition", then video generation is a "comprehensive ability test".

Why is it the most difficult? Because it not only needs to process the content of each frame but also ensure the coherence between frames, the smoothness of movements, and even consider factors such as audio - video synchronization and camera movement.

Currently, the leader in this field is Google. Their Veo 3 model has an ELO value as high as 1247, making it one of the most powerful video - generation models at present.

However, it is not invincible.

China's Kuaishou launched Kling 2.0, and Alibaba also released the Wan 2.1 version. Their ELO values reached 1053 and 1039 respectively. Although they are still not as good as Veo 3, they already have good video - generation capabilities, especially in scenarios such as short - video production and advertising creativity.

More importantly, these models are developed based on China's local application needs and have stronger support for Chinese prompts and local scenarios.

Seeing this, you may ask: Is China fully committed to multi - modal development? Yes, and it is using a combined approach.

Large technology companies have started to integrate language, image, and video capabilities into the same platform.

For example, Alibaba's ModelScope not only provides text models but also starts to support image generation and video - editing functions. Huawei's Pangu Ultra series has also gradually added visual and voice modules.

Tencent's Hunyuan has directly launched a one - stop multi - modal interface, allowing developers to call multiple capabilities at once.

The advantages of this approach are obvious: instead of letting each modality operate independently, it enables them to work together to form a complete AI capability system.

Therefore, the development of Chinese AI is a systematic layout. This competition is not over yet. Judging from the current performance, China is becoming an important participant in global AI multi - modal innovation.

05

Even if China is advancing rapidly, is the direction correct? The AI competition between China and the United States has an impact that goes beyond technology itself and involves multiple aspects such as policy, industry, security, and ethics.

In fact, there is a basis and necessity for cooperation between the two sides in many aspects.

AI is a technological field that highly depends on openness and collaboration. If communication is completely cut off, it will not only affect the innovation speed but also lead to resource waste and redundant development.

Specifically, in which aspects can cooperation take place? There are three points.

Firstly, basic research needs to be shared and interconnected. Although there are currently some restrictions, there is still a lot of room for global cooperation in areas such as basic models, algorithm optimization, and evaluation methods.

For example, key technologies such as the Transformer architecture and diffusion models were initially promoted by researchers from multiple countries. If this spirit of openness can be continued, it will provide more possibilities for AI development.

Secondly, industry standards need to be jointly established.

Model evaluation standards, ethical norms, security mechanisms... all these require global consensus. Currently, multiple international organizations are promoting the construction of AI governance frameworks, including UNESCO, IEEE, OECD, etc. As the two major AI powers, both China and the United States are actually involved.

Although the paces and stances are different, it at least shows that everyone realizes that AI needs rules, and these rules need to be jointly formulated.

Thirdly, commercial interests drive cooperation. Although the government may sometimes restrict exports and technology transfers, the motivation for cooperation between enterprises always exists.

For example, some Chinese enterprises hope to use NVIDIA's chips for training, and American enterprises also hope to enter the Chinese market. This interdependent relationship objectively provides room for cooperation.

Of course, we must also face the reality: there is indeed an obvious trend of confrontation in the AI competition between China and the United States.