GuShiio.com DeepSeek’s open source initiative is expected to attract more cloud service providers, reduce the cost of self-built cloud computing centers or privatized deployment, form a stronger ecosystem, and enhance China’s global leadership in the field of artificial intelligence.
DeepSeek Open Source Week, a large-scale model developed by domestic artificial intelligence companies in depth, has come to an end.
On February 28, DeepSeek recently announced that it will open source the Fire-Flyer File System (3FS) propeller for DeepSeek’s full data access. According to reports, this is a parallel file system that leverages the full bandwidth of modern solid state drives (SSDs) and remote direct memory access (RDMA) networks to accelerate and promote all data access operations on the DeepSeek platform.
At this point, DeepSeek Open Source Week officially ended. Previously, at noon on February 21, DeepSeek announced that it would open source five code bases and share its research progress with the global developer community in a completely transparent manner. New content will be unlocked every day to further share new progress, and define this plan as “Open Source Week”.
“The five projects provided by this Open Source Week cover the core aspects of AI development-from hardware performance squeezing, model training optimization to speed up data processing. The goal is to allow developers to use it out of the box, reduce technical thresholds and costs, and enable Developers can use large models more efficiently, at a low cost, and extensively.” Regarding the specific impact of open source, Wang Wei, a professor at the School of Data Science and Engineering at East China Normal University, told The Paper that it is expected to attract more cloud service providers, reduce the cost of self-built cloud computing centers or privatize deployment, and form a stronger ecosystem., compete with other large model ecosystems.
He mentioned that the reason why DeepSeek chose the open source route and did not worry about being surpassed by its peers is because its core competitive model and profit sources do not need to rely on sales model services to make profits. In addition, it also shows that DeepSeek has considerable technical confidence and “believes that it will not be quickly overtaken by industry competitors and can persist in leading the continued development of the most advanced technologies under its own system.”
“In a larger sense, DeepSeek’s open source is guiding global standard setting and forming a stronger DeepSeek ecosystem. Through open source, developers from more countries can be attracted to join the DeepSeek ecosystem, which will greatly Improve China’s global leadership in the field of artificial intelligence.” Wang Wei believes that if DeepSeek forms a strong open source ecosystem on a global scale, it will promote domestic chip manufacturers to further adapt, have broader scenarios and markets, and achieve a closed business loop.
What projects does DeepSeek open source this time and what is the significance for the big model industry? The Paper reporter sorted out all open source projects from February 24 to 28. Since these projects involve many technical terms, the reporter used the DeepSeek web version to provide specific explanations of the projects:
On February 24, the first open source code library was FlashMLA.
FlashMLA is known as an “accelerator” that enhances the potential of graphics cards. FlashMLA is DeepSeek’s efficient MLA decoding core for Hopper GPUs. It is optimized for variable-length sequences and is now in production.
FlashMLA is specifically used to optimize the computing efficiency of graphics cards, especially NVIDIA’s GPUs. For example, when AI processes sentences of different lengths (such as long text and short text), it can dynamically allocate computing power, avoid waste of resources, and bring processing speeds close to the hardware limit. Actual measurements show that this makes AI translation, content generation and other tasks faster and more cost-effective.
DeepEP is known as the “communication steward” of large model training and is specially designed to improve the efficiency of large model training. For example, when multiple AI expert models (MoE architectures) work together, it can efficiently coordinate communication between them and reduce Delay and resource consumption, while supporting low-precision calculations (such as FP8), further saving computing power.
On February 26, DeepSeek announced the open source DeepGEMM: DeepGEMM is known as the “power-saving expert” of matrix computing. It is a tool to optimize matrix multiplication (the core calculation of AI training). Speed is improved through low-precision calculation (FP8), and then Nvidia CUDA technology is used to correct errors. It is fast and accurate. The code is only 300 lines. It is simple to install and suitable for rapid deployment.
On February 27, DeepSeek opened source two tools and a dataset: DualPipe, EPLB, and analytical data from training and reasoning frameworks. Liang Wenfeng himself was also among the developers.
DualPipe is mainly used to solve the “latency” problem in pipeline parallelism. For example, if multiple task steps have different speeds, it can be scheduled in both directions and reduce idle time. EPLB is used to automatically balance the GPU load. When some AI expert models have too heavy tasks, it will copy the tasks to idle graphics cards to avoid “busy ones die and idle ones die.”
On February 28, DeepSeek announced the open source 3FS (Fire-Flyer File System) system: 3FS is called the “extremely fast combination” of data processing. It adopts a distributed file system, utilizes high-speed storage and network technologies (such as SSD, RDMA), and achieves a data read speed of 6.6TB per second, suitable for massive data training.
Why on earth did DeepSeek choose open source? How will this open source radiate to the industry?
“DeepSeek’s open source is equivalent to building a floating bridge on Nvidia’s AI moat.” Tan Jian, associate professor of intelligent interaction design at Beijing University of Posts and Telecommunications, told the paper reporter that more importantly, these open source modules of the DeepsSeek team have proved their ability to deeply analyze the tight coupling pattern between NVIDIA CUDA and parallel computing chips. This is also considered by traditional AI research. It is an unshakable software and hardware infrastructure and Nvidia’s broad moat.
Tan Jian believes thatThis week’s intensive open source models and algorithms have reshaped the AI hardware operating logic. It not only effectively responds to the previous query that Deepseek training model still requires huge computing power, but also predicts that the open source of these core libraries will greatly stimulate the global AI software and hardware team. The innovative vitality of the team。
Regarding the impact of DeepSeek on AI and even the chip industry, Tan Jian said: On the one hand, the AI model software research team can reduce hardware requirements through algorithm optimization (such as low-rank attention compression). On the other hand, algorithm optimization exposes existing AI chips. Design flaws, my country’s AI chip research and development team can also learn from these algorithms to rewrite the design of internal computing units and communication buses. It is expected that domestic software and hardware integrated AI models may be used in various sub-fields in the future, opening a new era in which hundreds of boats compete for the application of AI models in my country.
Liu Cong, a senior observer of the big model industry, told reporters that DeepSeek’s wave of open source can be called the “conscience of the industry.” Although it has limited direct application value for ordinary users and most users, it is very useful for practitioners of underlying technology.
Liu Cong believes that DeepSeek has provided open source for all the infra (infrastructure) optimizations mentioned in the DeepSeek-V3 paper, and many open source frameworks can add these optimization strategies.As hardware resources are compressed again, there may be a wave of API price cuts, thus continuing to lead the industry in openness and transparency.
As a big open source model, the popularity of DeepSeek has driven open source to become a new trend in today’s big models. Baidu and Ali have announced open source for their big models, and open source for leading manufacturers seems to have become a common choice.
Shen Xiangyang, Chairman of the Board of Governors of the Hong Kong University of Science and Technology and a foreign academician of the National Academy of Engineering, said at the 2025 Global Developer Pioneers Conference (GDC) that although the current share of closed source still exceeds the share of open source, it will change dramatically in the next year or two, balancing open source and closed source to lead the future. “In the era of big models, open source is not as much and as fast as before. I think through Shanghai’s efforts, I believe that open source will get better and better. The team in China and the team in Shanghai will definitely lead the trend of open source.”
“Although open source for large models seems to have become mainstream in China, it has not become uniform globally.” Wang Wei said frankly that for example, OpenAI, a head model manufacturer, still maintains a closed-source approach, and even DeepSeek has reservations in the open source process. For example, it has not implemented open source in the training data and training process.
“There is a route dispute between open source and closed source, which is not only between enterprises, but may even rise to the national level.In the digital economy era, the cost of information copying is almost zero. DeepSeek’s choice of open source can quickly occupy the market and gain a large amount of monthly income. After that, it may consider adopting other business models to make profits. However, the traditional closed-source model needs to invest a lot of advertising costs to occupy the market and promote users.”
It is worth noting that the competition for artificial intelligence research and development is becoming increasingly fierce. On February 27, local time, artificial intelligence giant OpenAI on the other side of the ocean released GPT-4.5 (research preview), calling it the company’s largest and most powerful chat model to date.
However, due to continued high investment and high cost, GPT-4.5 is facing huge controversy. Public information shows that developers can call GPT-4.5 directly in the API, but the input token is 30 times more expensive than GPT-4o, and the output token is 15 times more expensive. Ultraman, CEO of OpenAI, said that although he wanted to launch GPT-4.5 Plus and Pro versions at the same time, the GPUs were exhausted, and tens of thousands of GPUs would be added next week, and then they would be launched to the Plus level.
What do you think of OpenAI’s newly launched GPT-4.5? Wang Wei believes that this reflects the two different development paths of closed source and open source. GPT-4.5 has great advantages in many evaluation capabilities, but it consumes huge computing power and money.”From our perspective, although it has advantages, it requires huge costs. We are more optimistic about sustainable development models like DeepSeek.”
In contrast, DeepSeek continues to lower costs and cost-effectiveness routes. On February 26, DeepSeek issued a price reduction notice: the peak period between 00:30 and 08:30 Beijing time every day is off-peak, and the price of API calls has been significantly reduced. Among them, DeepSeek-V3 has been reduced to 50% of the original price, and DeepSeek-R1 has been reduced to 25%.
