cstr/firered-ocr-crispembed-GGUF

FireRed-OCR — CrispEmbed GGUF

GGUF conversion of FireRedTeam/FireRed-OCR for CrispEmbed.

FireRed-OCR is a Qwen3-VL 2B fine-tune with Format-Constrained GRPO training for tables, formulas, and structured document OCR. 92.94% on OmniDocBench v1.5.

Architecture

• Vision: Qwen3-VL ViT (24 layers, 1024d, patch 16, deepstack [5,11,17])
• LLM: Qwen3-2B (28 layers, 2048d, GQA 16/8, QK norms, mRoPE interleaved)
• Training: Format-Constrained GRPO with rewards for formula syntax, table integrity, hierarchical closure

Models

File	Quant	Size	Notes
firered-ocr-f16.gguf	F16	4.0 GB	Full precision
firered-ocr-q8_0.gguf	Q8_0	2.2 GB	Best quality/size
firered-ocr-q4_k.gguf	Q4_K	1.6 GB	Smallest
firered-ocr-ref.gguf	F32	17 MB	Reference activations for parity testing

Usage

crispembed --ocr firered-ocr-q8_0.gguf document.png

Runs on the qwen2vl_ocr engine (auto-detected from GGUF architecture qwen3vl).

License

Apache-2.0 (FireRedTeam/FireRed-OCR)

Provenance and EU AI Act Art. 53 note

• Upstream model: FireRedTeam/FireRed-OCR — published by FireRedTeam.
• Upstream licence: apache-2.0. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
• What was done here: format conversion and/or quantisation only (GGUF). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
• Training data: documented — where it is documented at all — by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository. No training-content summary was found on the upstream model card at the time of writing; that documentation gap is upstream's and is not filled here.
• Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.