FireRed-OCR-GGUF

FireRed-OCR-GGUF

FireRed-OCR from FireRedTeam is a specialized framework that transforms general Large Vision-Language Models into pixel-precise structural document parsing experts, tackling "Structural Hallucination" issues like disordered rows and invented formulas through a shift to "structural engineering" paradigms, achieving SOTA 92.94% on OmniDocBench v1.5—vastly outperforming DeepSeek-OCR 2, OCRVerse, and giants like Gemini-3.0 Pro or Qwen3-VL-235B. Its key innovations include Format-Constrained GRPO (Group Relative Policy Optimization) for enforcing syntactic validity (no unclosed tables or invalid LaTeX), a "Geometry + Semantics" data factory with geometric clustering and multi-dimensional tagging for balanced long-tail layouts, and a progressive pipeline: multi-task pre-alignment for spatial grounding, specialized SFT for standardized full-image Markdown output, and GRPO self-correction via RL. Demonstrating in-the-wild robustness on FireRedBench complex layouts over traditional systems like PaddleOCR, it excels in high-fidelity parsing of tables, equations, forms, and multi-column documents for real-world automation.

Model Files

Quants Usage

(sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants)

Here is a handy graph by ikawrakow comparing some lower-quality quant types (lower is better):