Roadmap

Status: ✅ Done | 🚧 WIP | 🔜 Soon | ❌ N/A

---

Text Extraction Models

Model	Params	PyTorch	VLLM	MLX	API
Qwen3-VL	2-32B	✅	✅	✅	✅
DotsOCR	1.7B	✅	✅	❌	✅
Nanonets OCR2	3B	✅	✅	✅	❌
Granite-Docling	258M	🔜	🔜	❌	❌
MinerU VL	1.2B	🔜	🔜	🔜	❌
LightOnOCR-2	1B	🔜	🔜	❌	❌
Chandra	9B	🔜	🔜	❌	❌
olmOCR-2	7B	🔜	🔜	❌	🔜
DeepSeek-OCR-2	3B	🔜	🔜	🔜	🔜

---

Layout Analysis Models

Model	Params	PyTorch	VLLM	MLX	API
DocLayoutYOLO	-	✅	❌	❌	❌
RT-DETR	-	✅	❌	❌	❌
Qwen Layout	2-32B	✅	✅	✅	✅
Surya Layout	-	🔜	❌	❌	❌
Florence-2	-	🔜	❌	❌	🔜

---

OCR Models

Model	Params	PyTorch	VLLM	MLX	API
Tesseract	-	✅	❌	❌	❌
EasyOCR	-	✅	❌	❌	❌
PaddleOCR	-	✅	❌	❌	❌
Surya OCR	-	🔜	❌	❌	❌
GOT-OCR2	700M	🔜	❌	❌	❌

---

Table Extraction Models

Model	Params	PyTorch	VLLM	MLX	API
TableFormer	-	✅	❌	❌	❌
TableTransformer	-	🔜	❌	❌	❌
Surya-Table	-	🔜	❌	❌	❌

---

Reading Order Models

Model	Params	PyTorch	VLLM	MLX	API
Rule-based (R-tree)	-	✅	❌	❌	❌

---

Tasks

Task	Status
Document Loading	✅
Text Extraction	✅
Layout Analysis	✅
OCR Extraction	✅
Table Extraction	✅
Reading Order	✅
Math Recognition	🔜
Structured Output	🔜
Chart Understanding	🔜

---

Infrastructure

Component	Status
Document class	✅
Pydantic configs	✅
Multi-backend	✅
Batch processing	✅
Modal deployment	✅
Testing framework	✅

---

Scripts Ready (Pending Integration)

These models have working scripts but aren't yet integrated into OmniDocs:

Model	Task	Scripts
Granite Docling	Text	VLLM, HF, MLX
MinerU VL	Text	VLLM, MLX

---

Under Consideration

Models being evaluated for future integration:

High Priority

Model	Use Case	Why
Marker	Full pipeline	Uses Surya, good tables
Granite Vision 3.3	Document understanding	IBM, good charts
Surya	OCR + Layout	Multi-language, modern

Specialized Use Cases

Use Case	Models Under Review
Handwriting	TrOCR, Surya
Scientific Papers	Nougat, Marker
Asian Languages	PaddleOCR-VL (109 langs)
Edge/Mobile	Granite-Docling (258M)
Forms & Receipts	DeepSeek-OCR-2

Framework Integration

Framework	Status
deepdoctection	🔜
Docling	🔜
Marker	🔜

---

Benchmarks Reference

OlmOCR-Bench (Higher = Better)

Model	Score	Params
LightOnOCR-2	83.2	1B
Chandra	83.1	9B
olmOCR-2	82.4	7B
DotsOCR	79.1	1.7B
Nanonets OCR2	~78	3B
DeepSeek-OCR	75.4	3B

Speed (Pages/Second on H100)

Model	Speed
LightOnOCR-2	5.7
PaddleOCR-VL	3.3
DeepSeek-OCR	2.3

---

Next Up

1. 🔜 Granite Docling - Integration from scripts
2. 🔜 MinerU VL - Integration from scripts
3. 🔜 Math Recognition - UniMERNet or Qwen
4. 🔜 Surya - Multi-language OCR + Layout

---

For detailed model specs, see ROADMAP_DETAILED.md.

Last Updated: February 2026