The View | How DeepSeek’s open-source breakthrough is reshaping AI innovation

The breakthrough performances of DeepSeek V3 and R1 do not guarantee a sustained edge for China’s artificial intelligence development, but they do highlight that the competitive advantages of US-based market leaders are less insurmountable than once believed. In the highly competitive AI landscape, where innovation cycles are compressed into months, top large language model (LLM) rankings can reshuffle with each new generation. These dynamic races now include Chinese LLMs competing at the highest level, challenging traditional US dominance.

On the first day of Lunar New Year, Alibaba unveiled Qwen 2.5 Max, claiming performance superiority over both DeepSeek V3 and leading US-based LLMs. While DeepSeek R1 has achieved parity with OpenAI’s o1, the newly released o3 offers enhanced capabilities.

DeepSeek has achieved impressive things with limited resources. However, maintaining competitive parity with US market leaders will require continuous improvements with expanded access to external resources, particularly through leveraging growing open source AI ecosystems.

What astounded the world about DeepSeek was not so much its strong performance but how it reached the heights of the industry with far less in terms of investments, computing power and time. This feat was accomplished through engineering optimisation and building upon existing foundations.

The company employed distillation techniques, where knowledge from larger, more complex models is transferred to smaller ones while maintaining robust performance. Such mutual learning, including distillation, is a common practice across industrial and academic AI development.