DeepSeek, a Chinese AI start-up, has announced its commitment to open-source technology by making five of its code repositories public next week. This move aims to foster transparency and collaboration within the AI community. The Hangzhou-based start-up, established in 2023, has gained attention amid the US-China competition in artificial intelligence. DeepSeek’s open-source approach contrasts with the closed models of companies like OpenAI and Anthropic. The company believes that sharing code will accelerate the journey of AI development, promoting community-driven innovation. DeepSeek’s commitment to open source has been critical to its rise, allowing free use and modification of its V3 and R1 models.

DeepSeek’s flagship DeepSeek-R1 model, launched in January 2025, has demonstrated reasoning and mathematical skills comparable to leading rivals. Unlike OpenAI’s ChatGPT and Anthropic’s Claude, DeepSeek is open source, allowing anyone to download, copy, and build upon it. Its code and technical explanations are freely shared, enabling global developers and organizations to access, modify, and implement. While Meta and Google offer models for viewing, their usage is restricted by licenses and undisclosed training data sets, unlike DeepSeek. Venture capitalist Marc Andreessen praised DeepSeek as “one of the most amazing and impressive breakthroughs I’ve ever seen – and as open source, a profound gift to the world”.

DeepSeek-V3 is a Mixture-of-Experts (MoE) language model with 671B total parameters, of which 37B are activated per token. It achieves a breakthrough in inference speed over previous models and tops the leaderboard among open-source models5. DeepSeek-V3 excels in English, code, math, and Chinese language tasks. The company offers free access to DeepSeek-V3 via its app and API platform. DeepSeek also provides various AI models, including DeepSeek R1, DeepSeek Coder V2, DeepSeek VL, DeepSeek Coder, and DeepSeek Math.

DeepSeek’s open-source nature allows anyone to build on top of its models. DeepSeek has made its models’ code publicly available, reinforcing its commitment to open source. The company’s DeepSeek-R1-Zero model, trained via reinforcement learning without supervised fine-tuning, demonstrated reasoning performance. DeepSeek-R1 incorporates cold-start data before reinforcement learning to address issues like repetition and poor readability. DeepSeek’s approach enables reasoning in a long context, potentially bootstrapped from scratch without extensive training on internet data.