Community gains from open source of Deepseek.
The startup of Chinese AI Deepseek has announcement Its intention to share technology behind its internal inference engine, a decision to improve collaboration within the Open Source community.
The framework of the engine and the company’s inference training played an essential role in accelerating performance and deployment of its models, in particular Deepseek-V3 and R1.
Built on Pytorch, the Deepseek training framework is supplemented by a modified version of the VLLM inference engine originally developed in the WE in UC Berkeley.
Although the company will not publish the complete source code of its engine, it will contribute its design improvements and will select the components as autonomous libraries.
These efforts are part of the broader Deepseek initiative, which started earlier this year with the partial release of its IA model code.
Despite this contribution, the Deepseek models are not below the standards of the open source initiative, because the training data and the complete framework remain limited.
The company has cited resources and limited infrastructure constraints such as reasons not to make the engine fully open. However, this decision was greeted as a significant gesture towards transparency and the sharing of knowledge in the AI sector.
Do you want to know more about AI, technology and digital diplomacy? If necessary, Ask our Chatbot Diplo!