Despite previous planning, Chinese company DeepSeek has yet to set an exact date for the release of its DeepSeek R2 model. It appears that the company’s CEO is still not satisfied with the model’s performance.
According to Reuters, DeepSeek CEO Liang Wenfeng believes that the performance of the R2 model is not yet satisfactory, which has led to a halt in its release process. R2, the anticipated successor to the successful reasoning model DeepSeek R1, was initially scheduled for release in May. The goal behind developing this new model was to generate better code, enhance reasoning capabilities, and support languages beyond English.
Delay in the Release of DeepSeek R2
In recent months, DeepSeek engineers have been working on refining and improving the R2 model, awaiting the final green light from Liang Wenfeng. However, this process has faced significant challenges — particularly the shortage of Nvidia server chips in China due to U.S. export restrictions, which has slowed down the implementation of the new model. This has been confirmed by The Information, citing staff from major Chinese cloud service providers that deliver DeepSeek’s models to enterprise clients.
A sudden surge in demand for the R2 model could place considerable strain on the infrastructure of Chinese cloud providers, as these companies rely on advanced Nvidia chips to run AI models. New export restrictions introduced by the Trump administration in April of this year have prevented the sale of Nvidia’s H20 chips to the Chinese market. At the time, these chips were the only AI processors that could still legally be exported to China.
Meanwhile, DeepSeek has been in contact with several major Chinese cloud companies and has shared technical specifications of the model to help host and distribute it via their servers. Currently, most cloud customers using the R1 model are running it on Nvidia’s H20 chips.
DeepSeek has not yet responded to Reuters’ request for comment, but it is clear that the success and release of the R2 model depend on several factors, including performance improvements and the state of the hardware market.
