GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Geong
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Geong
Model Tree
dots-studio/dots.ocr
Input Modalities
Output Modalities
Supported Functionality
DARL introduces two components on top of a vision-language backbone:
At inference, multiple tokens are generated in parallel via special initialization tokens, then verified incrementally from left to right by the sliding window.
| Base model | dots.ocr |
| Architecture | DotsOCRForCausalLM (dots_ocr) |
| Parameters | ~3B (bfloat16) |
| Language model | 28 layers, hidden size 1536, 12 heads (2 KV heads) |
| Vision encoder | dots_vit, 42 layers, patch size 14 |
| Context length | 131072 |
| Precision | bfloat16 |
Requires trust_remote_code=True since the model ships custom modeling code.
Note: This repo ships the weights, config and tokenizer only. The image preprocessor config is not included — load the processor from the base model dots-studio/dots.ocr as shown below, or copy
preprocessor_config.jsonfrom there into this repo.
python
from transformers import AutoModelForCausalLM, AutoProcessorfrom PIL import Image model = AutoModelForCausalLM.from_pretrained( "your-org/DARL", trust_remote_code=True, torch_dtype="bfloat16", attn_implementation="flash_attention_2", device_map="cuda",)# processor comes from the base model, see note aboveprocessor = AutoProcessor.from_pretrained("dots-studio/dots.ocr", trust_remote_code=True) image = Image.open("document.jpg")prompt = "Recognize image as Markdown format" messages = [{"role": "user", "content": [ {"type": "image", "image": image}, {"type": "text", "text": prompt},]}]text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)inputs = processor(text=[text], images=[image], return_tensors="pt").to(model.device) outputs = model.generate(**inputs, max_new_tokens=4096)print(processor.batch_decode(outputs[:, inputs.input_ids.shape[1]:], skip_special_tokens=True)[0])transformers >= 4.51.3Evaluated on document parsing benchmarks:
Speedup scales with model size, up to 2.78× on 14B backbones. See the paper and repository for full numbers.
bibtex
@inproceedings{darl2026, title={DARL: Efficient Document-to-Markup Generation via Look-Ahead Diffusion Trajectory Sampling}, author={Yang, Wentao and Shi, Yongxin and Tang, Rui and Zhang, Peirong and Wu, Shihang and He, Huiguo and Huang, Zheng and Peng, Dezhi and Liao, Minghui and Jin, Lianwen}, booktitle={European Conference on Computer Vision (ECCV)}, year={2026}}Built on top of dots.ocr and Qwen-VL. Supported by the National Natural Science Foundation of China (62476093) and the Natural Science Foundation of Guangdong Province (2026A1515012038).
Apache 2.0
from transformers import AutoModelForCausalLM, AutoProcessorfrom PIL import Image model = AutoModelForCausalLM.from_pretrained( "your-org/DARL", trust_remote_code=True, torch_dtype="bfloat16", attn_implementation="flash_attention_2", device_map="cuda",)# processor comes from the base model, see note aboveprocessor = AutoProcessor.from_pretrained("dots-studio/dots.ocr", trust_remote_code=True) image = Image.open("document.jpg")prompt = "Recognize image as Markdown format" messages = [{"role": "user", "content": [ {"type": "image", "image": image}, {"type": "text", "text": prompt},]}]text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)inputs = processor(text=[text], images=[image], return_tensors="pt").to(model.device) outputs = model.generate(**inputs, max_new_tokens=4096)print(processor.batch_decode(outputs[:, inputs.input_ids.shape[1]:], skip_special_tokens=True)[0])@inproceedings{darl2026, title={DARL: Efficient Document-to-Markup Generation via Look-Ahead Diffusion Trajectory Sampling}, author={Yang, Wentao and Shi, Yongxin and Tang, Rui and Zhang, Peirong and Wu, Shihang and He, Huiguo and Huang, Zheng and Peng, Dezhi and Liao, Minghui and Jin, Lianwen}, booktitle={European Conference on Computer Vision (ECCV)}, year={2026}}