Is DeepSeek’s intellectual property protection sufficient to counter potential risks?

In January 2025, DeepSeek, a Chinese AI startup, released its large language model, DeepSeek-R1. According to DeepSeek’s disclosures, this model achieves performance comparable to that of ChatGPT’s o1 model across multiple benchmark tests, while significantly reducing training costs. Beyond significantly improving efficiency, DeepSeek has taken the bold step of open-sourcing its model, permitting users to freely use, modify, and commercialise it. For an open-source AI company such as DeepSeek, safeguarding its innovations presents a unique challenge.

This article will analyse DeepSeek’s current intellectual property protection strategy and the risks it must address in the future.

Starting from February 2025, DeepSeek declared that it would open-source five core codebases. DeepSeek has already open-sourced several distillation models and partial reasoning data sets for mathematics, science, coding, and puzzles.

On one hand, DeepSeek’s open-source data provides global developers and researchers with greater opportunities to explore how to achieve large-model capabilities with smaller-scale resources. On the other hand, this policy broadens DeepSeek’s brand reach and helps to build a powerful community ecosystem. These advantages position DeepSeek as a strong competitor to closed-source models and a significant force in the future AI landscape.

However, it is important to note that DeepSeek’s open-source plan does not disclose the following:

• The complete training data set and training code;

• The technical details of data processing and storage; and

• The experimental details and training strategies.

These core competitive elements remain protected as trade secrets. Thus, DeepSeek’s current open-source plan can also be regarded as an ‘open weight’ project with trade secret protection.

In addition to protecting its core intellectual property as trade secrets, DeepSeek has also laid out patents for its technological innovations. The main entity of DeepSeek – Hangzhou DeepSeek Artificial Intelligence Co., Ltd. – was established in July 2023. Since the company is relatively new, no patent applications have been found under its name as of February 2025. However, considering that DeepSeek’s founder, Wenfeng Liang, controls multiple AI-related companies, such as Hangzhou DeepSeek and Ningbo High-Flyer Quant, it is reasonable to expand the search to analyse the patent portfolios of these DeepSeek-related enterprises.

Patent searches reveal that DeepSeek and its related enterprises began filing patent applications in 2019. As of February 2025, they have submitted 19 invention patent applications, with seven granted. These inventions cover technologies such as cluster training management methods, data communication methods, and data storage methods, partially reflecting DeepSeek’s innovations in efficiency optimisation. However, it is notable that all the current patent applications are filed exclusively in China. Based on publicly available information, it remains unclear whether DeepSeek has pursued international patent protection for its innovations.

Based on the above analysis, DeepSeek has adopted a collaborative protection mechanism combining trade secrets and patents for its innovations, which is a strategy commonly used by open-sourcing companies. However, given the potential multibillion-dollar global market DeepSeek may face, whether its current strategy can adequately address future risks and challenges remain questionable, particularly in terms of balancing open-source collaboration with robust intellectual property protection.

Despite DeepSeek’s non-disclosure of trade secrets such as training system code, complete training framework, and data processing toolchain, the rapidly evolving AI landscape has made reverse engineering increasingly feasible. Techniques such as knowledge distillation, reinforcement learning, and multi-stage training enable researchers to approximate and replicate proprietary technologies. For instance, researchers at Hugging Face Inc. have stated that they have launched the Open-R1 project, aiming to reverse engineer the DeepSeek-R1 inference model to create a fully open-source replica of the R1 model. The growing computational power further facilitates such reverse engineering efforts, compelling companies to be more cautious in choosing open-source content.

DeepSeek must also carefully identify which innovations to open-source in the future; for example, prioritising trade secret protection for technologies with a lower possibility to be reverse engineered, while utilising patent protection for innovations with a higher likelihood of being reverse engineered.

Based on the current disclosure, DeepSeek’s patent applications only cover a subset of its innovations. While it has filed patents related to cluster management and scheduling, its core technical breakthroughs – such as a sparse mixture of experts and multi-head latent attention hybrid architecture, a multi-token prediction mechanism, and a group relative policy optimisation algorithm for reinforcement learning – remain unpatented.

Furthermore, DeepSeek’s patent strategy is geographically limited, with no traceable overseas patent filings to date. This phenomenon aligns closely with the intellectual property risks exposed by the trademark squatting case it encountered in the US in January 2025.

The above scenarios highlight insufficient planning in DeepSeek’s intellectual property protection strategy across scope‌, category, and geographical dimensions.

DeepSeek currently lacks a sufficient number of patents to build a moat for its intellectual property rights. By the end of 2024, the global count of generative AI patents exceeded 90,000, with leading companies such as Tencent and Samsung each holding over 1,000 generative AI patents.

The number of patents directly reflect a company’s technological accumulation and R&D strength, and an extensive patent portfolio is essential for establishing AI technology barriers. For open-source companies such as DeepSeek, proactive patent filings are crucial to prevent third-party squatting on similar technologies, thereby mitigating litigation risks and gaining negotiation leverage. However, DeepSeek’s patent portfolio remains underdeveloped, necessitating accelerated filings to catch up with industry benchmarks.

Although DeepSeek currently adopts a collaborative protection mechanism combining trade secrets and patents for its innovations, the increasing feasibility of reserve engineering in the AI field suggests that the role of trade secret protection will become more limited. For DeepSeek, strengthening patent protection of its innovative works has become increasingly important.

The battle for patent protection as a core innovation safeguard has only just begun, and DeepSeek must act swiftly to fortify its intellectual property strategy.