The Chinese AI company with the Wallogo is causing a stir in the markets. Just over two weeks after its release, the Deepseek app has replaced ChatGPT as the top-rated free AI application in the App Store. This success has significant impacts on the stock market, especially in the chip sector.
The stock prices of chip manufacturers like Nvidia and the Dutch company ASML, which supplies almost all semiconductor manufacturers with machines for chip production, have dropped significantly. In Silicon Valley, Deepseek is also causing a stir. According to research from The Information, the work of the Chinese startup has caused some excitement in Meta’s AI department.
Deepseek can compete with OpenAI while saving resources. Deepseek-R1, the latest AI model from the Chinese startup, not only competes with GPT-3 but even surpasses the OpenAI model in some benchmarks. For example, the Chinese model performs better in a series of tests that evaluate the mathematical abilities of such models. It also holds its own against the American competition in generating program code and complex chains of reasoning.
The real highlight is the various methods Deepseek uses to conserve hardware resources during the training and execution of its models while improving performance. This includes the Mixture-of-Experts architecture (MoE). This technique ensures that only the necessary parameters are used to complete a task. Instead of using all 671 billion parameters, only 37 billion are employed during execution. This requires significantly less hardware resources and still delivers convincing results.
Another Deepseek innovation is the Multi-Head Latent Attention (MLA), which allows AI models to focus on multiple aspects of the input data to better understand them. This mechanism is considered the main reason for the excellent performance of the AI model Deepseek-V3, introduced at the end of 2024. Additionally, Deepseek has developed various methods to transfer the knowledge of an AI model into a smaller and thus more efficient model. This way, comparatively strong models can be created that require significantly less computing power.
Deepseek embraces Open Source. Although OpenAI’s company name suggests a certain openness, this has long been the case. Details about proprietary AI models remain sparse, which has repeatedly drawn criticism in the past, for example from Meta’s AI chief Yann LeCun or OpenAI co-founder Elon Musk. Deepseek, on the other hand, has released its latest models under an open-source license. The AI model platform Hugging Face has already launched the project Open R1 based on Deepseek-R1. This aims to rebuild the entire development pipeline based on the components published by Deepseek. This will ultimately benefit both research and companies alike.
Deepseek delivers technically impressive AI models but, as a Chinese company, also censors certain questions on its own platform. While the underlying models are freely available, users should not forget that Deepseek is a Chinese company and must comply with Beijing’s censorship regulations. For example, if you ask about the Tiananmen Massacre on the Deepseek website or app, you won’t get an answer. Instead, the chatbot explains that Deepseek, as a “responsible company,” is committed to Chinese laws and “the social morality of China.”