CoolFace
Modelpublic

cstr/internvl2-1b-crispembed-GGUF

sourceHugging Facemitupdated 2mo agoView on Hugging Face
0likes160downloads
Model Card

InternVL2-1B — CrispEmbed GGUF

GGUF conversions of OpenGVLab/InternVL2-1B for use with CrispEmbed.

Smallest competitive VLM for OCR — ideal for edge, mobile, and WASM deployment.

Model Details

PropertyValue
ArchitectureInternVL2 (InternViT-300M + Qwen2-0.5B)
Total Parameters~0.9B
Vision EncoderInternViT-300M-448px (24L, 1024d, identical to InternVL2.5-2B)
ProjectorPixel unshuffle (4:1) + LayerNorm + Linear + GELU + Linear
LLM DecoderQwen2-0.5B-Instruct (24L, 896d, GQA 14/2, SwiGLU, RMSNorm)
Input Resolution448x448 per tile, dynamic tiling (1-12 tiles)
LicenseMIT
OCRBench779

Available Quantizations

FileSizeCompressionNotes
internvl2-1b-f16.gguf2.3 GB1xFull precision
internvl2-1b-q8_0.gguf955 MB2.4xGood quality
internvl2-1b-q4_k.gguf~600 MB~4xSmallest, vision Q8_0 floor

Parity Verification

All components verified against Python reference (cos=1.000000):

  • —Vision encoder: 4/4 layers PASS
  • —Projector: PASS
  • —LLM decoder (Qwen2): 2/2 layers PASS

Credits

Provenance and EU AI Act Art. 53 note

  • —Upstream model: OpenGVLab/InternVL2-1B — published by OpenGVLab.
  • —Upstream licence: mit. 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.