PaddleOCR-VL-1.6 — CrispEmbed GGUF

Files

Quantization quality. This is a small 0.9B model, so 4-bit quantization is noticeably lossy. In testing, q8_0 transcribes fully and matches the fp16 reference, while q4_k tends to terminate early (e.g. it emitted only "The quick brown fox jumps" and stopped, dropping the rest of the line) — a flattened output distribution from 4-bit weights, not a bug. Behavior is identical on CPU and Metal. Use q8_0 for accurate OCR; reach for q4_k only when memory is tight and truncated/approximate output is acceptable.

Model

• Architecture: NaViT-style ViT (27L, 1152d) + ERNIE-4.5-0.3B LLM (18L, 1024d, 16/2 GQA, MRoPE, SwiGLU)
• Parameters: ~0.9B (same architecture as PaddleOCR-VL-0.9B, improved training)
• OmniDocBench: 96.3% (SOTA)
• Languages: 109+ (multilingual)
• License: Apache 2.0

Usage

License

Provenance and EU AI Act Art. 53 note

• Upstream model: PaddlePaddle/PaddleOCR-VL-1.6 — published by PaddlePaddle.
• 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.