Infinity-Parser2-Flash-GGUF

Infinity-Parser2-Flash is a low-latency document understanding model from infly-ai, one of two variants in the Infinity-Parser2 flagship family (alongside the accuracy-optimized Infinity-Parser2-Pro), engineered for fast inference while consolidating robust multi-modal parsing into a unified architecture trained via an upgraded synthetic data engine spanning nearly 5 million diverse document samples and a novel multi-task reinforcement learning approach with verifiable rewards across document parsing, element parsing, chart parsing, chemical formula parsing, document VQA, and general multimodal understanding. It delivers a 3.68x speedup over the previous Infinity-Parser-7B model (increasing throughput from 441 to 1,624 tokens/sec) while still posting strong benchmark results — 86.0% on olmOCR-Bench, 72.2% on ParseBench, and 91.98% on OmniDocBench-v1.6 — outperforming frontier models like DeepSeek-OCR-2 and MinerU2.5 on several document-parsing tasks, though trailing its larger Pro sibling on layout analysis, chart/chemical formula parsing, and general multimodal benchmarks (e.g., MMMU, AI2D, MathVista). It extracts structured layout with bounding boxes, category labels, and per-element text (LaTeX for formulas, HTML for tables, Markdown for text), supports command-line and Python API usage via the infinity_parser2 package with vLLM, transformers, or vLLM-server backends, and is released under Apache-2.0, with known limitations primarily around English/Chinese-only support, degraded accuracy on complex charts and rotated table elements, and no fine-grained text formatting (bold, italic, strikethrough) capture.

Model Files

llama.cpp

LLM inference in C/C++ — https://github.com/ggml-org/llama.cpp