dots.ocr — CrispEmbed GGUF
GGUF conversion of rednote-hilab/dots.ocr for CrispEmbed.
dots.ocr unifies layout detection, text extraction, table parsing, and formula recognition in a single VLM. 100+ languages. 88.4% on OmniDocBench.
Architecture
- Vision: Custom ViT (42 layers, 1536d, patch 14, 2D RoPE, SwiGLU FFN, PatchMerger 2x2)
- LLM: Qwen2 (28 layers, 1536d, GQA 12/2, standard RoPE, attention_bias=true)
- Training: Prompt-based task switching (OCR, layout, table, formula)
Models
| File | Quant | Size |
|---|---|---|
| dots-ocr-f16.gguf | F16 | 2.8 GB |
| dots-ocr-q8_0.gguf | Q8_0 | 1.5 GB |
| dots-ocr-q4_k.gguf | Q4_K | 912 MB |
Usage
License
MIT (rednote-hilab/dots.ocr)
- Downloads last month
- -
Hardware compatibility
Log In to add your hardware
8-bit
16-bit
Model tree for cstr/dots-ocr-crispembed-GGUF
Base model
rednote-hilab/dots.ocr