GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
ProKope-AI
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
ProKope-AI
Model Tree
Input Modalities
Output Modalities
Supported Functionality
ProKope-4B-Eagle-Eye is a 4.2B-parameter multimodal vision-language model developed by ProKope AI for document parsing, structured HTML table reconstruction (rowspan/colspan), mathematical LaTeX transcription, multi-column reading-order extraction, and chart interpretation.
The model uses the Qwen3_5ForConditionalGeneration architecture (4.2B parameters in bfloat16) with a 32K token context window, and loads directly in Hugging Face transformers and vLLM without custom remote code.
Evaluated with deterministic greedy decoding (do_sample=False, 0-shot Pass@1) across standard document OCR and table extraction benchmarks:
| Benchmark | Dataset | Split | Metric | ProKope-4B-Eagle-Eye | LightOnOCR-3-4B | Chandra-OCR-2 (4B) | olmOCR-2-7B |
|---|---|---|---|---|---|---|---|
| PulseBench-Tab | pulse-ai/PulseBench-Tab | train (1,820) | TEDS Similarity | 89.40% | 84.40% | 78.90% | 70.50% |
| olmOCR-Bench | allenai/olmOCR-bench | test (7 splits) | Pass Rate | 87.06% | 86.30% | 85.80% | 82.40% |
| ParseBench | llamaindex/ParseBench | test (5 pillars) | Macro Average | 86.01% | 75.10% | 70.10% | 43.30% |
| fr-bench-pdf2md | pulsia/fr-bench-pdf2md | long_table | Unit-Test Pass Rate | 85.11% | 64.00% | 72.20% | 61.40% |
| Real5-OmniDocBench | PaddlePaddle/Real5-OmniDocBench | test (5 scenarios) | Overall Similarity | 76.61% | — | — | — |
allenai/olmOCR-bench Category Breakdown| Category | ProKope-4B-Eagle-Eye | LightOnOCR-3-4B | Chandra-OCR-2 (4B) | olmOCR-2-7B |
|---|---|---|---|---|
ArXiv Math (arxiv_math) | 95.56% | 91.10% | 86.90% | 83.00% |
Old Scans Math (old_scans_math) | 93.10% | 89.10% | 89.10% | 82.30% |
Multi-Column (multi_column) | 92.86% | 85.60% | 82.10% | 83.70% |
Old Scans (old_scans) | 85.19% | 52.50% | 51.10% | 47.70% |
Tables (table_tests) | 85.07% | 90.20% | 92.10% | 84.90% |
Long Tiny Text (long_tiny_text) | 79.41% | 93.20% | 93.70% | 81.90% |
Headers & Footers (headers_footers) | 78.26% | 89.20% | 91.40% | 96.10% |
| Overall | 87.06% | 86.30% | 85.80% | 82.40% |
llamaindex/ParseBench Pillar Breakdown| Pillar | Metric | ProKope-4B-Eagle-Eye | LightOnOCR-3-4B | Nemotron-3.5-Parse |
|---|---|---|---|---|
Table (table) | TEDS Similarity | 93.95% | 83.80% | 84.67% |
Layout (layout) | Reading Order & Content | 90.46% | 68.30% | 70.67% |
Text Content (text_content) | 12-Rule Verbatim Match | 87.98% | 89.90% | 87.96% |
Text Formatting (text_formatting) | 11-Rule Hierarchy & Emphasis | 82.51% | 67.60% | 68.28% |
Chart (chart) | Label & Value Extraction | 75.17% | 66.10% | 64.40% |
Overall (mean) | 5-Pillar Macro Average | 86.01% | 75.10% | 75.20% |
| Property | Value |
|---|---|
| Architecture | Qwen3_5ForConditionalGeneration |
| Total Tensor Footprint | 4.2B combined parameters (9.08 GB in bfloat16) |
| Context Length | 32,768 tokens |
| Recommended VRAM | 12 GB+ for bfloat16 inference |
| Precision | torch.bfloat16 |
| License | Apache 2.0 (Weights gated under manual review) |
transformers)Requires transformers, torch, and pillow:
python
import torchfrom PIL import Image, ImageOpsfrom transformers import AutoProcessor, Qwen3_5ForConditionalGeneration model_id = "ProKope-AI/ProKope-4B-Eagle-Eye" processor = AutoProcessor.from_pretrained(model_id)model = Qwen3_5ForConditionalGeneration.from_pretrained( model_id, torch_dtype=torch.bfloat16, device_map="auto",).eval() image = Image.open("document_page.png").convert("RGB")image = ImageOps.autocontrast(image, cutoff=1).resize((1024, 768), Image.Resampling.BILINEAR) messages = [ { "role": "user", "content": [ {"type": "image", "image": image}, { "type": "text", "text": "[Visual Geometry: 1024x768px | Domain: complex_grid_table_html]\nExtract all tables and text on this document page into canonical HTML `<table>` structure and Markdown.", }, ], }] prompt = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)inputs = processor(text=[prompt], images=[image], return_tensors="pt").to(model.device)inputs["pixel_values"] = inputs["pixel_values"].to(torch.bfloat16) with torch.no_grad(): generated_ids = model.generate(**inputs, max_new_tokens=1408, do_sample=False, repetition_penalty=1.08) response = processor.decode(generated_ids[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)print(response)vLLMbash
vllm serve ProKope-AI/ProKope-4B-Eagle-Eye \ --dtype bfloat16 \ --max-model-len 16384 \ --limit-mm-per-prompt image=2Access to ProKope-4B-Eagle-Eye weights is gated (manual) for benchmark verification, academic research, and commercial licensing evaluation. Production deployment is available through two options:
POST /v1/parse): Pay-as-you-go multi-page PDF and image extraction returning clean Markdown, HTML tables, and structured JSON without managing GPU infrastructure.vLLM / FastAPI) for private VPC, HIPAA, SOC 2, and air-gapped on-premise deployments on a single 16GB–24GB GPU.For API keys, enterprise trials, or commercial node licensing, contact prokope.model@gmail.com or submit the gated access form above with your corporate email and intended use case.
rowspan/colspan), mathematical LaTeX transcription, and chart extraction.ParseBench [chart]: 75.17%), sub-6pt footnote text (olmOCR-bench [long_tiny_text]: 79.41%), or complex running watermarks (olmOCR-bench [headers_footers]: 78.26%) may contain transcription errors. Verify critical financial or medical outputs independently.bibtex
@misc{canto2026prokope4beagleeye, author = {Alan Canto and Wesley Foreman}, title = {ProKope-4B-Eagle-Eye: Multimodal Document Parsing and Table Reconstruction Model}, howpublished = {Hugging Face}, year = {2026}, url = {https://huggingface.co/ProKope-AI/ProKope-4B-Eagle-Eye}}import torchfrom PIL import Image, ImageOpsfrom transformers import AutoProcessor, Qwen3_5ForConditionalGeneration model_id = "ProKope-AI/ProKope-4B-Eagle-Eye" processor = AutoProcessor.from_pretrained(model_id)model = Qwen3_5ForConditionalGeneration.from_pretrained( model_id, torch_dtype=torch.bfloat16, device_map="auto",).eval() image = Image.open("document_page.png").convert("RGB")image = ImageOps.autocontrast(image, cutoff=1).resize((1024, 768), Image.Resampling.BILINEAR) messages = [ { "role": "user", "content": [ {"type": "image", "image": image}, { "type": "text", "text": "[Visual Geometry: 1024x768px | Domain: complex_grid_table_html]\nExtract all tables and text on this document page into canonical HTML `<table>` structure and Markdown.", }, ], }] prompt = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)inputs = processor(text=[prompt], images=[image], return_tensors="pt").to(model.device)inputs["pixel_values"] = inputs["pixel_values"].to(torch.bfloat16) with torch.no_grad(): generated_ids = model.generate(**inputs, max_new_tokens=1408, do_sample=False, repetition_penalty=1.08) response = processor.decode(generated_ids[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)print(response)vllm serve ProKope-AI/ProKope-4B-Eagle-Eye \ --dtype bfloat16 \ --max-model-len 16384 \ --limit-mm-per-prompt image=2@misc{canto2026prokope4beagleeye, author = {Alan Canto and Wesley Foreman}, title = {ProKope-4B-Eagle-Eye: Multimodal Document Parsing and Table Reconstruction Model}, howpublished = {Hugging Face}, year = {2026}, url = {https://huggingface.co/ProKope-AI/ProKope-4B-Eagle-Eye}}