README.md

metadata

license: gemma
base_model:
  - Cognitive-Lab/NetraEmbed
pipeline_tag: visual-document-retrieval
library_name: transformers
tags:
  - text-generation-inference
language:
  - en
  - es
  - fr
  - de
  - it
  - hi
  - mr
  - sa
  - kn
  - te
  - ta
  - ml
  - zh
  - ja
  - ko
  - ar
  - bn
  - gu
  - or
  - pa
  - ru
  - th

NetraEmbed-GGUF

NetraEmbed from Cognitive-Lab is a state-of-the-art multilingual multimodal embedding model powered by a Gemma3-4B-IT backbone with SigLIP vision encoder, designed for visual document retrieval via BiEncoder architecture that encodes images of documents and text queries into compact single dense vectors supporting Matryoshka dimensions of 768 (fastest, 95% accuracy retention), 1536 (balanced), or 2560 (maximum accuracy) for flexible inference without model reloading. It achieves groundbreaking performance on Nayana-IR Bench (22 languages) with 0.716 NDCG@5 on cross-lingual tasks—152% improvement over ColPali-v1.3—and 0.738 on monolingual, while being 250x more storage-efficient (~10KB per document vs. 2.5MB multi-vector) than traditional approaches, preserving visual elements like charts, tables, and layouts without OCR errors. Ideal for scalable semantic search across millions of multilingual PDFs/scans using cosine similarity in vector DBs like FAISS, Milvus, or Pinecone, it enables enterprise-grade cross-lingual document discovery for revenue charts, hierarchies, or diagrams in diverse scripts.

NetraEmbed [GGUF]

File Name	Quant Type	File Size	File Link
NetraEmbed.BF16.gguf	BF16	7.77 GB	Download
NetraEmbed.F16.gguf	F16	7.77 GB	Download
NetraEmbed.F32.gguf	F32	15.5 GB	Download
NetraEmbed.Q8_0.gguf	Q8_0	4.13 GB	Download
NetraEmbed.mmproj-bf16.gguf	mmproj-bf16	851 MB	Download
NetraEmbed.mmproj-f16.gguf	mmproj-f16	851 MB	Download
NetraEmbed.mmproj-f32.gguf	mmproj-f32	1.67 GB	Download
NetraEmbed.mmproj-q8_0.gguf	mmproj-q8_0	591 MB	Download

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):