The Convergence of Chinese Open Source
In early February, when the Chinese open-source large model DeepSeek topped application market download charts in 140 countries and regions worldwide, OpenAI publicly accused DeepSeek of using distilled data from ChatGPT without permission.
Rather than salvaging OpenAI’s reputation, this accusation was met with widespread ridicule from researchers worldwide.
Now, another contender, fully embracing the ‘distillation’ buff, has emerged.
On April 13th, Kunlun Wanwei launched the Skywork-OR1 (Open Reasoner 1) series models, outperforming Alibaba’s Qwen-32B in the same scale and aligning with DeepSeek-R1.
How can Kunlun Wanwei, a company with limited financial resources, create a SOTA-level large model? The official explanation is that their models are based on DeepSeek-R1-Distill-Qwen-7B and DeepSeek-R1-Distill-Qwen-32B.
As the names suggest, DeepSeek’s models distilled Alibaba’s Qwen series models.
While leveraging excellent open-source models, Kunlun Wanwei is also contributing to the open-source community. Unlike DeepSeek, which only open-sourced model weights, Kunlun Wanwei has also open-sourced its datasets and training code, aligning more closely with the concept of ‘true open source.’ This means that any user can attempt to replicate their model training process.
Kunlun Wanwei’s achievement demonstrates the most important aspect of open source: it not only provides users with a free and readily available product but also enables more developers to stand on the shoulders of predecessors, rapidly and cost-effectively advancing technology.
In fact, amid last year’s industry discussions about the bottleneck in large model pre-training, the pace of iteration of Chinese large models has accelerated this year, with more and more companies investing in open source.
Alibaba Cloud’s Tongyi Qianwen open-sourced its new visual model Qwen2.5-VL on Chinese New Year’s Eve and released and open-sourced its new reasoning model Qwen-32B in early March, topping the trending list of the global mainstream AI open-source community Hugging Face on the day it was open-sourced.
Stepwise then open-sourced three multimodal large models in about a month, the latest of which is the image-to-video model Step-Video-TI2V, which supports the generation of videos with controllable motion amplitude and lens movement, and also comes with certain special effects generation capabilities.
Zhipu announced in April that it would open-source the 32B/9B series GLM models, covering base, reasoning, and contemplative models, all under the MIT license agreement.
Even Baidu, which was once closed source, announced that it would fully open-source the Wenxin large model on June 30.
Compared to the growing prosperity of the domestic open-source ecosystem, American large model companies still mainly focus on closed source, which has given Chinese large models a rare opportunity to go overseas. DeepSeek has allowed Indonesian education company Ruangguru to optimize its teaching model at a low cost; Singaporean B2B travel technology company Atlas has integrated Qwen into its intelligent customer service system to achieve 24/7 multilingual support.
Why Closed Source in the US, Open Source in China?
The tendency towards closed source in the US AI industry and the increasing openness of Chinese AI are the inevitable results of the different AI development environments in the two countries.
The US AI industry is mainly led by tech giants and VCs (venture capitalists), who have huge expectations for capital returns from AI. Therefore, US AI model companies generally have a strong belief in technology, that is, to pursue technological leadership, achieve a certain degree of market monopoly, and then create huge profits, and their ecosystem is naturally inclined to closed source.
Taking OpenAI’s development history as an example, it started as a non-profit entity during its establishment, but has since become increasingly closed off. GPT-1 was fully open source, GPT-2 was partially open source and encountered opposition before being fully open source, GPT-3 officially went closed source, and then GPT-4 further strengthened the closed-source strategy, with model architecture and training data completely confidential, and even restricting the API calling frequency of corporate users.
Although OpenAI said that closing source is based on compliance and controlling the abuse of technology, the market generally believes that the landmark event of OpenAI’s shift to closed source was its reaching a hundred-billion-dollar cooperation with Microsoft, embedding GPT-3 into Azure cloud services to form a ‘technology-capital’ closed loop.
When Microsoft first disclosed its investment in OpenAI in its financial report in October last year, it said: ‘We have invested in OpenAIGlobal, LLC, with a total investment commitment of $13 billion, and the investment is accounted for using the equity method.’
The so-called equity method can also be understood as that Microsoft’s investment in OpenAI is aimed at obtaining returns rather than pure charitable research. Obviously, OpenAI’s selling high-priced APIs through a closed-source ecosystem is its current largest source of revenue, and has become the biggest obstacle to OpenAI’s unwillingness to open source.
Anthropic, which was founded from OpenAI’s ‘split,’ has been determined to take the closed-source route from the beginning, and its large model product Claude has fully adopted the closed-source model.
Even META’s Llama, the only open-source leader in the United States, added two anti-friend clauses when open-sourcing:
- Open-source models cannot be used for products and services with more than 700 million monthly active users before META approves them.
- The output content of Llama models cannot be used to train and improve other large language models.
It can be seen that even for open-source models, Meta’s core purpose is still to build its own AI ecosystem rather than technical inclusiveness.
The United States has chosen an AI strategy based on closed source with open source as a supplement at the capital level, which can be said to be purely commercial considerations. In contrast, China’s top-down top-level design has attached importance to open source from the beginning, reflecting an industry-first path under the concept of independent control.
As early as 2017, the Chinese government released the ‘New Generation Artificial Intelligence Development Plan,’ which clearly proposed to accelerate the deep integration of AI with the economy and society, and deploy to build the first-mover advantage of AI development. In 2021, open-source-related content was explicitly included in China’s ‘14th Five-Year Plan,’ which triggered active promotion of technological innovation by local governments.
Mei Hong, an academician of the Chinese Academy of Sciences, once said that the future development of language models must rely on open-source platforms. Only in an open environment can the security and trustworthiness of data uploads and business integration for users in various industries be ensured.
The ‘Special Action Plan for Digital Empowerment of Small and Medium-sized Enterprises (2025-2027)’ issued by the Ministry of Industry and Information Technology and other four departments in December last year clearly supports the Open Atom Open Source Foundation to establish a ‘Small and Medium-sized Enterprise AI Open Source Special Project’ to provide reproducible and easy-to-promote training frameworks, testing tools, and other resources to lower the technical threshold for enterprises.
A more realistic problem is that due to the potential technological blockade by the United States, China cannot simply be a follower in the AI field, but must build an independent domestic ecosystem. Building another closed-source ecosystem under the ecosystem that the United States has already built with closed source as the main focus is tantamount to building a car behind closed doors. Only an open-source ecosystem can quickly help the development of the AI industry.
In addition to top-level support, various local governments have also made real money investments in the open-source ecosystem.
The Z Fund, jointly established by Zhipu and Beijing State-owned Assets, which focuses on large model ecosystem investment, announced that it would invest 300 million yuan to support the development of the AI open-source community worldwide. Any startup project based on open-source models (not limited to Zhipu open-source models) can apply.
The divergence between China and the United States in their open-source and closed-source strategies for the AI industry is essentially a fundamental difference in development logic. The United States is driven by capital, and the profit-seeking demands of tech giants and VCs have spawned a closed-source ecosystem of ‘technology monopoly-high-priced realization.’ Even if Meta tries to open source, it is difficult to escape the shackles of commercial barriers. China relies on top-level design, with ‘technology equity + industrial collaboration’ as its core concept, and builds an open ecosystem through policy empowerment, making open source a basic infrastructure for lowering technical thresholds and promoting the integration of the real economy. This strategic choice not only shapes the different paths of the AI industries in the two countries but also heralds the acceleration of the global AI ecosystem from ‘monopoly competition’ to ‘open and win-win.’
Good Enough is Good Enough
China’s AI open-source ecosystem is not only accelerating the AI industrialization development in China and the world but also putting the United States’ belief in technology first into an awkward trap.
Faced with the increasing pressure brought by the DeepSeek effect, Meta released Llama4 on April 5, claiming it to be the strongest multimodal large model in history.
However, after actual testing, this is a model that is disappointing. The context length of 10m tokens often goes wrong, the initial ball test is difficult to complete, and the 9.11 > 9.9 comparison size error occurs. Within a few days of the model’s launch, scandals such as executive resignations and test cheating were also confirmed by internal employees.
More news proves that Llama4 can be said to be a product that Zuckerberg rushed to put on the shelves. So the question is, why did Zuckerberg have to launch it in April?
As mentioned earlier, the US AI industry has a confusing belief in technology, believing that their products must be the strongest and most advanced, so they have started an arms race. However, the diminishing marginal effect of training AI has caused large manufacturers to consume huge amounts of costs, and not only has the technical threshold not been built, but they have fallen into the quagmire of computing power bottlenecks.
After OpenAI released GPT-4o’s image generation function, Altman tweeted a few days later that their GPUs were ‘burning out.’ Less than a week after Gemini2.5 was released, the head of GoogleAIStudio said that they were still plagued by ‘rate limits,’ and developers could only send 20 requests per minute. It seems that no company can cope with the inference needs of super-large models.
In fact, the United States is falling into a misunderstanding. The person in charge of the Zhiyuan Research Institute said: ‘If a new model uses 100 times the cost to run out of a 10-point benchmark score increase, then this new model is meaningless for more than 80% of application scenarios because there is no cost performance.’
Chinese large model companies are accelerating the open-source ecosystem. They seem to no longer be competing for the top spot, but instead have won more customers, especially industrial customers, with their ‘good enough’ approach.
Compared to the tens of millions of budgets for government and enterprise customers, many companies and institutions have urgent AI needs but do not have so many existing solutions. Using open-source models to develop their own solutions has almost become their only choice:
- Baosteel uses the ‘large model + small model’ for key metallurgical engineering processes for intelligent early warning of production equipment.
- China Coal Science and Industry Group’s ‘Coal Science Guardian Large Model ChinamjGPT’ reduces equipment downtime and maintenance costs by 30% and 20%, respectively.
- Shanghai Mengbo Intelligent Internet of Things Technology has created an edge-cutting detection and continuous annealing furnace process optimization application platform based on a lightweight large model.
- Mifei Technology has realized intelligent prediction, maintenance, and management of automated material handling systems in semiconductor wafer fabs based on large model technology.
These are all representative cases of open-source models being implemented in industrial scenarios.
In addition to industrial uses, the open-source ecosystem can also help more public welfare undertakings.
The Shanshui Nature Conservation Center is committed to the protection of snow leopards and plateau ecosystems. The infrared cameras it sets up take a large number of photos or videos every quarter. It is extremely inefficient and time-consuming to rely on manual identification of snow leopard traces. Huawei Ascend is cooperating with the Shanshui Nature Conservation Center to identify snow leopard traces. Huawei has open-sourced the relevant models and tools for infrared image species recognition in Sanjiangyuan, lowering the threshold for participating in AI development and allowing more research and protection institutions using the model to benefit. People can work together to optimize the model in terms of datasets, data processing, and data cleaning.
The ‘Bazaar’ Effect of Open Source
Eric Raymond, the flag-bearer of the open-source software movement, proposed a metaphor in his 1999 book ‘The Cathedral and the Bazaar’: The traditional, closed-source software development model is like building a cathedral. The software is carefully designed and built by a few experts (architects) in an isolated environment and is only released to users after it is finally completed; The open-source development model is like a bustling, seemingly chaotic but vibrant bazaar. Software development is open, decentralized, and evolutionary.
The book believes that for many types of software projects, especially complex system-level software (such as operating system kernels), the open, collaborative, and decentralized ‘bazaar’ development model, although it may seem chaotic, is actually more efficient, produces higher quality, and more robust software than the traditional, closed, and centralized ‘cathedral’ model. It can discover and fix errors faster and better absorb user feedback and community contributions through mechanisms such as ‘release early, release often’ and leveraging large-scale peer review (‘enough eyeballs’), thereby promoting rapid iteration and innovation of software.
The huge success of open-source projects such as Linux has verified Raymond’s point.
The open-source movement has brought the United States and the world a huge value far exceeding its own investment. A 2024 research report from Harvard University stated: ‘Open-source invested $4.15 billion and created $8.8 trillion invalue for companies (that is, every $1 invested creates $2,000 in value). Without open source, corporate spending on software would be 3.5 times what it is now.’
Today, Chinese companies have learned this. American AI companies seem to have forgotten it.
In fact, for Chinese large model companies, even if they do not consider social benefits, choosing to embrace the open-source ecosystem is not unprofitable for the companies themselves.
Many large model companies have told Observer.com that open source does not mean giving up commercialization. Open source still has the profit logic of open source. Compared to whether it is open source or not, how to better serve customers technically is the key issue.
Taking Zhipu AI as an example, it claims to be the only company in China that fully benchmarks OpenAI, but compared to OpenAI’s closed-source strategy, it is one of the most determined practitioners of open-source strategy in the industry.
Zhipu took the lead in open-sourcing China’s first Chat large model ChatGLM-6B in 2023. Since its establishment nearly six years ago, Zhipu has open-sourced more than 55 models, with a cumulative download volume of nearly 40 million times in the international open-source community.
Zhipu told Observer.com that Zhipu hopes that its open-source strategy will contribute to building Beijing into a ‘global open-source capital’ for artificial intelligence.
Specifically, at the commercial level, Zhipu chose to attract a developer ecosystem through open source and provide paid customized solutions to B-end and G-end customers.
In addition to selling solutions, selling APIs is also an important profit link.
Taking DeepSeek as an example, the first business of the open-source model is the sale of high-performance APIs. Although basic services are free, companies can provide high-performance API services and charge based on usage. The API pricing for DeepSeek-R1 is 1 yuan per million input tokens and 16 yuan per million output tokens. If the free token quota is used up or the basic API cannot meet the needs, users tend to use the paid version to maintain the stability of business processes.
Compared to companies that only have model services, Alibaba has chosen another open-source monetization model: ecosystem bundling.
Alibaba’s Qwen series, as an open-source pioneer, attracts developers to use cloud computing and other infrastructure through full-modal open source, forming a closed-loop scenario. Their model is just an introduction in the early stage, and the goods with marked prices are actually cloud services.
The globalization application of Chinese open-source large models has shifted from ‘technology following’ to ‘ecosystem dominance.’ When the United States is caught in the dilemma of ‘closed-source monopoly’ and ‘open-source out of control,’ China is reconstructing the underlying logic of the global AI open-source ecosystem through ‘agreement innovation + scenario cultivation.’ The ultimate battlefield of this game is not in the competition of parameter scale but in the trillion-dollar market of deep integration of AI technology and the real economy.