gemma
3B parameters
Commercial OK
Reviewed May 2026

ColPali v1.3

3B-parameter visual document retriever built on PaliGemma-3B using a ColBERT-style late-interaction objective. Encodes a PDF page as a grid of patch embeddings, skipping OCR/layout parsing entirely. Sets SOTA on the ViDoRe benchmark for visual document retrieval.

License: mit·Context: 0 tokens
BLK · VERDICT

Our verdict

OP · Fredoline Eruo|VERIFIED MAY 29, 2026
unrated

The clear winner for visual-document RAG when the corpus is PDF-heavy and full of charts/tables. Budget for the storage overhead of multi-vector indexes and pick your vector store deliberately — that's the only real friction.

Overview

3B-parameter visual document retriever built on PaliGemma-3B using a ColBERT-style late-interaction objective. Encodes a PDF page as a grid of patch embeddings, skipping OCR/layout parsing entirely. Sets SOTA on the ViDoRe benchmark for visual document retrieval.

Strengths

  • SOTA on ViDoRe (~81 nDCG@5 on v1.x) — beats unstructured.io + BGE pipelines by 15+ points
  • Skips OCR entirely — handles charts, tables, handwriting, low-quality scans natively
  • Late-interaction (ColBERT-style) gives per-token precision while staying queryable
  • MIT licensed, PaliGemma backbone — fully Apache/MIT downstream
  • Indexes 1 page in <100ms on an RTX 3090

Weaknesses

  • Multi-vector index: ~1030 embeddings per page — storage cost is ~100x dense retrievers
  • Needs a late-interaction-capable store (Vespa, Qdrant multi-vector, plaid-x) — not a drop-in for FAISS
  • Inherits PaliGemma's Gemma license restrictions on the base weights (LoRA adapter is MIT)
  • Only the retriever — pair with a generative VLM (Qwen2.5-VL, GPT-4o, etc.) for answers

Quantization variants

Each quantization trades model quality for file size and VRAM. Q4_K_M is the most popular starting point.

QuantizationFile sizeVRAM required
Q4_K_M1.7 GB3 GB

Get the model

HuggingFace

Original weights

huggingface.co/vidore/colpali-v1.3

Source repository — direct quantization required.

Hardware that runs this

Cards with enough VRAM for at least one quantization of ColPali v1.3.

Compare alternatives

Models worth comparing

Same parameter band, plus what's one tier above and below — so you can decide what actually fits your hardware.

Frequently asked

What's the minimum VRAM to run ColPali v1.3?

3GB of VRAM is enough to run ColPali v1.3 at the Q4_K_M quantization (file size 1.7 GB). Higher-quality quantizations need more.

Can I use ColPali v1.3 commercially?

Yes — ColPali v1.3 ships under the mit, which permits commercial use. Always read the license text before deployment.

What's the context length of ColPali v1.3?

ColPali v1.3 supports a context window of 0 tokens (about 0K).

Source: huggingface.co/vidore/colpali-v1.3

Reviewed by RunLocalAI Editorial. See our editorial policy for how we research and verify model claims.

Related — keep moving

Before you buy

Verify ColPali v1.3 runs on your specific hardware before committing money.