GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
enclavelabs
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
enclavelabs
Model Tree
allenai/olmOCR-2-7B-1025
Input Modalities
Output Modalities
Supported Functionality
Given an image containing Devanagari text (Hindi, Marathi, Sanskrit, Nepali, Pali), returns the Unicode transcription. Unlike iter-3 which was trained on word crops only, iter-4 also handles full document pages without falling into generation loops.
For full-page PDFs, use it via the EnclaveScribe agent pipeline which handles page rasterization and generation-config.
Tested against a 6-page Gazette of India Extraordinary notification (dense Hindi + English) with the runtime bad_words_ids workaround disabled:
| Signal | iter-3 | iter-4 |
|---|---|---|
| Total chars extracted | 7,122 | 13,646 |
| Pages dead-looped | 2/6 | 0/6 |
| Max page chars (of 4,096 ceiling) | 2,430 | 4,077 |
| Wallclock (6 pages, g5.xlarge) | 20 min | 12 min |
Iter-3 pages 4 and 5 exhausted the full generation budget emitting <tool_call> blocks and produced empty output after the post-hoc regex strip. Iter-4 completes every page inside the budget with real document text.
Iter-4 is also more faithful to source script conventions — for example, it preserves Arabic numerals inside English sections (where iter-3 hallucinates Devanagari digits like २११३ instead of 2113).
See the full head-to-head in reports/iter4/GAZETTE_TEST.md.
python
import torchfrom transformers import AutoProcessor, AutoModelForImageTextToTextfrom peft import PeftModelfrom PIL import Image BASE = "allenai/olmOCR-2-7B-1025"ADAPTER = "enclavelabs/enclave-scribe-devanagari-iter4" processor = AutoProcessor.from_pretrained(BASE)model = AutoModelForImageTextToText.from_pretrained( BASE, dtype=torch.bfloat16, device_map="auto")model = PeftModel.from_pretrained(model, ADAPTER)model.eval() image = Image.open("hindi_page.png").convert("RGB")messages = [{ "role": "user", "content": [ {"type": "image", "image": image}, {"type": "text", "text": "Extract this document page as clean markdown. " "Preserve the original script exactly."}, ],}]text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)inputs = processor(text=[text], images=[image], return_tensors="pt").to(model.device) with torch.no_grad(): out = model.generate( **inputs, max_new_tokens=4096, do_sample=False, repetition_penalty=1.1, # belt-and-suspenders; iter-4 no longer strictly needs it ) print(processor.batch_decode( out[:, inputs.input_ids.shape[1]:], skip_special_tokens=True)[0].strip())Note: iter-4 no longer dead-loops on the pages that broke iter-3, but we recommend keeping repetition_penalty=1.1 in production. It costs nothing and stays as insurance.
allenai/olmOCR-2-7B-1025enclavelabs/enclave-scribe-devanagari), loaded via resume_adapter so training continues rather than starting freshallenai/olmOCR-2-32B-1025) with the same corpus for more LoRA capacitytests/fixtures/pdfs/gazette_moef_2024_06_07.pdf before shippingFollow github.com/Enclave-Labs-Inc/enclave-scribe for updates.
text
@software{enclavescribe_devanagari_iter4_2026, title = {EnclaveScribe: Self-hostable Indic OCR — Devanagari adapter (iter-4)}, author = {Enclave Labs}, year = {2026}, url = {https://huggingface.co/enclavelabs/enclave-scribe-devanagari-iter4},}MIT — same as the base model and the EnclaveScribe repo.
@software{enclavescribe_devanagari_iter4_2026, title = {EnclaveScribe: Self-hostable Indic OCR — Devanagari adapter (iter-4)}, author = {Enclave Labs}, year = {2026}, url = {https://huggingface.co/enclavelabs/enclave-scribe-devanagari-iter4},}import torchfrom transformers import AutoProcessor, AutoModelForImageTextToTextfrom peft import PeftModelfrom PIL import Image BASE = "allenai/olmOCR-2-7B-1025"ADAPTER = "enclavelabs/enclave-scribe-devanagari-iter4" processor = AutoProcessor.from_pretrained(BASE)model = AutoModelForImageTextToText.from_pretrained( BASE, dtype=torch.bfloat16, device_map="auto")model = PeftModel.from_pretrained(model, ADAPTER)model.eval() image = Image.open("hindi_page.png").convert("RGB")messages = [{ "role": "user", "content": [ {"type": "image", "image": image}, {"type": "text", "text": "Extract this document page as clean markdown. " "Preserve the original script exactly."}, ],}]text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)inputs = processor(text=[text], images=[image], return_tensors="pt").to(model.device) with torch.no_grad(): out = model.generate( **inputs, max_new_tokens=4096, do_sample=False, repetition_penalty=1.1, # belt-and-suspenders; iter-4 no longer strictly needs it ) print(processor.batch_decode( out[:, inputs.input_ids.shape[1]:], skip_special_tokens=True)[0].strip())