← All models

dots.vlm1 Instruct

RedNote (Xiaohongshu) · 2025-08-06 · 672.3B parameters

Dots Studio's released dots.vlm1 instruction checkpoint combines a vision encoder trained from scratch with the DeepSeek V3 language model for general multimodal understanding and generation. Its original model card reports broad visual reasoning, OCR and document benchmarks; this artifact is separate from dots3-note Preview. Released on August 6, 2025, its checkpoint contains 672,323,119,104 parameters including the vision encoder and projector and excluding FP8 quantization-scale tensors. The supported context limit and native reasoning controls are not established by the reviewed card; architecture fields describe its language backbone, and the tokenizer assigns 128,858 distinct IDs rather than the padded 129,280 embedding rows.

Benchmark scores