GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
cloudaocr
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
cloudaocr
Model Tree
amad-iq/amad-vlm6
Input Modalities
Output Modalities
Supported Functionality
Clouda OCR is an open-source, Arabic-first OCR and Document AI project. Clouda OCR Stage-1 is a vision-language model for image-to-text document processing, built on the Qwen2.5-VL family (qwen2_5_vl) through the AMAD base model. It targets Arabic OCR, scanned-document transcription, and document-AI research and experimentation.
The release is Stage-1 and experimental. It is not presented as a production-ready system.
The completed benchmark contains 462 pages: 62 KITAB-R pages and 400 MISRAJ pages. All 462 pages completed successfully, with no failed pages and no duplicate page IDs. Metrics were recalculated from all 462 successful per-page records.
| Metric | Result |
|---|---|
| CER | 3.744885% |
| Normalized CER | 3.739644% |
| WER | 4.429895% |
| Normalized WER | 4.426912% |
| Pages | 462 |
| Failed | 0 |
| Duplicates | 0 |
| Dataset composition | Pages |
|---|---|
| KITAB-R | 62 |
| MISRAJ | 400 |
Benchmark timing is intentionally not reported as an official result because the completed timing run was contended and non-canonical. No base-model improvement percentage is published: a directly comparable pre-training/base-model benchmark is not currently available.
This release is the result of the completed Stage-1 training run:
| Item | Value |
|---|---|
| Method | Supervised fine-tuning (SFT) with LoRA |
| Verified training volume | 10,000 verified training sample-passes |
| Available training pool | 56,136-record GOLD_PASS training pool |
| Final Stage-1 loss | 6.6155 |
The 10,000 figure is a count of training sample-passes, not a confirmed count of unique pages. Other training hyperparameters are omitted because they are not verified by the public release artifacts.
Suitable use cases include:
This Stage-1 release has not yet been fully validated for every handwriting style, Arabic script or font, historical manuscript, table or document-layout type, or production-critical and high-stakes document-processing workflow.
Benchmark coverage is finite, and performance can vary with scan quality, typography, degradation, layout complexity, and document domain. Further Stage-2 training and broader evaluation are planned.
The example below follows the repository's verified qwen2_5_vl architecture (Qwen2_5_VLForConditionalGeneration) and processor (Qwen2_5_VLProcessor, loaded through AutoProcessor). It uses a local image named page.png.
python
import reimport torchfrom transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration repo = "cloudaocr/clouda-ocr-amad-vlm6"model = Qwen2_5_VLForConditionalGeneration.from_pretrained( repo, dtype=torch.bfloat16, device_map="auto")processor = AutoProcessor.from_pretrained(repo) messages = [{"role": "user", "content": [ {"type": "image", "image": "page.png"}, {"type": "text", "text": "Extract the text in the image. Give only the final text."},]}]inputs = processor.apply_chat_template( messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt").to(model.device) output = model.generate(**inputs, max_new_tokens=4096, do_sample=False)text = processor.batch_decode( output[:, inputs["input_ids"].shape[1]:], skip_special_tokens=True)[0]print(re.sub(r"<think>.*?</think>", "", text, flags=re.S).strip())The model configuration and processor metadata were inspected to verify the model and processor classes. The repository also includes successful local-load and inference-smoke-test artifacts. Install a Transformers version compatible with the repository's transformers_version metadata before use.
The exact model revision used for the final benchmark was:
2043093ee04bc48a39132299f7d1b06b1a992939
Benchmark metrics were recalculated from all 462 successful records. Evaluation artifacts are preserved separately from this model repository. Public benchmark tooling or data may be linked later where licensing permits.
If you use this model, please cite the project:
bibtex
@software{clouda_ocr_stage1_2026, title = {Clouda OCR Stage-1}, author = {{Clouda OCR}}, year = {2026}, url = {https://huggingface.co/cloudaocr/clouda-ocr-amad-vlm6}}Clouda OCR Stage-1 is an experimental Stage-1 research release. Reported results refer only to the documented 462-page benchmark. No official speed claim is made from the completed benchmark run.
This model release is provided under the Apache License 2.0.
import reimport torchfrom transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration repo = "cloudaocr/clouda-ocr-amad-vlm6"model = Qwen2_5_VLForConditionalGeneration.from_pretrained( repo, dtype=torch.bfloat16, device_map="auto")processor = AutoProcessor.from_pretrained(repo) messages = [{"role": "user", "content": [ {"type": "image", "image": "page.png"}, {"type": "text", "text": "Extract the text in the image. Give only the final text."},]}]inputs = processor.apply_chat_template( messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt").to(model.device) output = model.generate(**inputs, max_new_tokens=4096, do_sample=False)text = processor.batch_decode( output[:, inputs["input_ids"].shape[1]:], skip_special_tokens=True)[0]print(re.sub(r"<think>.*?</think>", "", text, flags=re.S).strip())@software{clouda_ocr_stage1_2026, title = {Clouda OCR Stage-1}, author = {{Clouda OCR}}, year = {2026}, url = {https://huggingface.co/cloudaocr/clouda-ocr-amad-vlm6}}