All terms
Models & Products
DeepSeek V3
DeepSeek's downloadable model that activates only part of its huge network per request.
Definition
DeepSeek V3 is a large, downloadable language model from DeepSeek built as a mixture of experts: it has a huge number of internal settings, but only a small fraction switch on to handle any given request, which saves computation. It uses a memory-saving form of attention and was trained using low-precision math on a reportedly modest budget, reaching performance close to leading models of its time. It is the base the DeepSeek-R1 reasoning model was built from.