← All models

DeepSeek-VL2

DeepSeek · 2024-12-13 · 27.5B parameters

DeepSeek-VL2 is the largest model of the DeepSeek-VL2 family, released December 13, 2024: a SigLIP vision encoder with dynamic tiling feeding a DeepSeekMoE language model with MLA (30 layers, 72 routed and 2 shared experts, 6 routed active, sigmoid routing). Its released weights contain 27,480,134,248 parameters, about 4.5B activated per token by DeepSeek's rounded figure, with a 4,096-token context.

Benchmark scores