dots.vlm1

dots.vlm1 is the first vision-language model in RedNote HiLab's dots family, from the team behind the Xiaohongshu social platform. It pairs a 1.2-billion-parameter NaViT vision encoder, trained entirely from scratch with dynamic resolution support, with the DeepSeek V3 large language model as its language backbone.

The model was trained with heavy synthetic multimodal data - tables, charts, documents, and interleaved web content - to push visual perception and reasoning, particularly OCR. RedNote positions it as setting a new ceiling for open-source vision-language models while staying competitive on pure-text tasks.

Honest limits: the headline results are from RedNote's own evaluation suite, not independent benchmarks. The release date here is era-level (2025), flagged honestly. License details should be checked on the official repository and model card.