SOTA on OmniDocBench v1.6 via pure data engineering — no architecture changes. Expanded training from <10M to 65.5M samples.

0.9B model (0.4B encoder + 0.5B decoder) achieves #1 on OmniDocBench v1.5. Outperforms models 260x larger including GPT-5.2 and Gemini-3 Pro.

0.9B compact VLM with NaViT dynamic resolution. Best text and formula recognition scores despite minimal size. Also introduces Real5-OmniDocBench for real-world robustness.

1B-param end-to-end model achieving new top-tier OmniDocBench score with data-training co-design and structure-aware training.

Tencent 2.5B specialized VLM for document parsing. Second-highest among specialized models.

End-to-end unified OCR model. Wins 6/8 benchmarks in its class. Strong chart reasoning capabilities.

3-stage training: multi-task pre-alignment, specialized SFT, and GRPO RL with format constraints. Qwen3-VL backbone.

Financial-domain document parser with cross-page consolidation and cell-level referencing. Competitive on OmniDocBench while excelling on FinDocBench.

Alibaba 4B document parsing model. Strong table and formula recognition.

30B MoE (3B active) vision-language model. Top proprietary VLM tier on document parsing.

Tencent Hunyuan 1B OCR model. Excellent formula recognition (95.96 CDM).

Tiny 0.1B OCR model matching much larger models. Remarkable efficiency.

Commercial OCR API model. Claims world-leading accuracy with confidence scores and structured extraction endpoints.

Google Gemini 3 Flash on OmniDocBench v1.5. Strong generalist VLM.

4B document parsing model with reinforcement learning approach.

Google Gemini 3 Pro on OmniDocBench v1.5.

Top OmniDocBench 1.5 score. Slight improvement over Qwen 3.5 35B-A3B (89.3).

Score as cited in Kimi K3 blog (source: Anthropic official). Fable 5 hit fallbacks on 35% of tasks.

Score as cited in Kimi K3 blog (source: OpenAI official).

Moonshot Kimi K2.5 1T param model on document parsing.

DeepSeek 3B second-gen OCR model. Improves on original DeepSeek-OCR.

3B pro variant of MonkeyOCR with strong text recognition.

ByteDance 3B OCR model v2. Balanced performance across document types.

4B document parsing model. Consistent across text, tables, formulas.

3B OCR model from Xiaohongshu (rednote). Strong formula recognition.

Score as cited in Kimi K3 blog (source: Anthropic official).

Compact 1.2B pro variant. Good accuracy for size class.

Score as cited in Kimi K3 blog (source: OpenAI official).

OpenAI GPT-5.2 on OmniDocBench v1.5. Weaker than specialized OCR models.

Document understanding focus. Beats GPT-5 mini (77.0). Particular strength in Korean document understanding.