Dolphin-v2-f32-GGUF

ByteDance Dolphin-v2 is a 3B-parameter vision-language model built on Qwen2.5-VL-3B with Native Resolution Vision Transformer (NaViT) encoder and autoregressive decoder, designed as a universal document parsing solution via a document-type-aware two-stage architecture that classifies digital-born vs. photographed documents before applying hybrid strategies—element-wise parallel parsing for clean PDFs and holistic parsing for distorted scans. It supports 21 element categories (headings sec_0-5, paragraphs, formulas in LaTeX, HTML tables, indented code blocks, figures, lists, etc.) with absolute pixel coordinates for precise localization, achieving state-of-the-art OmniDocBench v1.5 scores of 89.45 overall (+14.78 over original Dolphin), 0.054 edit distance for text/reading order, 86.72% CDM for formulas, and 87.02/90.48 TEDS/TEDS-S for tables at 0.1729 FPS on 8-12GB VRAM GPUs. Specialized modules (P_formula, P_code, P_table, P_paragraph) enable structured JSON/Markdown/HTML outputs for privacy-focused local inference in healthcare/legal/finance, outperforming general VLMs in speed (2x faster) and accuracy across distortions, skews, and perspectives.

Dolphin-v2 [GGUF]

Quants Usage