← All models

DeepSeek-VL 1.3B Chat

DeepSeek · 2024-03-11 · 2B parameters

DeepSeek-VL 1.3B Chat is the small instruction-tuned vision-language model of the DeepSeek-VL release of March 2024, pairing a SigLIP vision encoder with a 1.3B DeepSeek LLM backbone for real-world image understanding, OCR and chart reading. Its released weights contain 1,975,235,584 parameters including the vision encoder; DeepSeek trained it with 4,096-token sequences although the configuration declares 16,384 positions.

Benchmark scores