deepseek-ai

DeepSeek's first model on the Causal Encoder-Decoder (CED) architecture — a sparse mixture-of-experts that activates 8B parameters on input and 16B on output from a 552B-parameter backbone, with image input and a 1M-token context window