DeepSeek-VL2

DeepSeek · 13 December 2024

MoE vision-language models with 1.0B, 2.8B and 4.5B active parameters, using MLA and dynamic tiling for high-resolution images.

Moves DeepSeek-VL onto DeepSeekMoE with Multi-head Latent Attention and dynamic-tiling vision encoding; strong on OCR, charts, documents and grounding for its active-parameter size.

Date

Friday, 13 December 2024

Lab

DeepSeek

Kind

open-weights

Access

open weights (restricted license)

Figures

Sources

1. arxiv.org/abs/2412.10302
2. github.com/deepseek-ai/DeepSeek-VL2

This record was checked against its sources on 6 October 2026. How we check

Related

• DeepSeek-VL (1.3B, 7B)