cstr/nanonets-ocr2-1.5b-crispembed-GGUF

Nanonets-OCR2-1.5B — CrispEmbed GGUF

Nanonets-OCR2-1.5B-exp (a pruned Qwen2-VL — 16 decoder layers instead of 28 — for document OCR, 12+ languages including German) converted to the single-file CrispEmbed GGUF layout, for the qwen2vl_ocr engine.

Converted from the upstream safetensors with models/convert-qwen2vl-to-gguf.py, then quantized with crispembed-quantize. CrispEmbed uses one combined file with its own tensor naming — these are not llama.cpp GGUFs and are not interchangeable.

File	Size	Notes
nanonets-ocr2-1.5b-q4_k.gguf	1346 MB	241 tensors quantized, 344 kept

Usage

crispembed -m nanonets-ocr2-1.5b --ocr document.png    # auto-downloads

Performance note

Full-page OCR pushes ~3200 vision patches through a 32-layer tower. On a busy machine that prefill is long — a first token can take many minutes if the CPU is contended. Give it a quiet machine before concluding it has hung.

Attribution & licence

Upstream model © Nanonets, Apache-2.0 — see nanonets/Nanonets-OCR2-1.5B-exp. Conversion and quantization do not relicense it. See CrispEmbed and its POLICY.md for intended purpose and acceptable use — OCR output is a probabilistic reconstruction, not a faithful copy, and VLM engines can confabulate through a smudge rather than leave it blank.

Provenance and EU AI Act Art. 53 note

• Upstream model: nanonets/Nanonets-OCR2-1.5B-exp — published by nanonets.
• 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.
• 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.