Best Visual Document Retrieval Models in 2026
Retrieve a PDF page by what it looks like, with no OCR step in front. We compare the late-interaction models on ViDoRe V3 accuracy, licence, and the index cost nobody puts on the model card: what a million pages actually costs to store.
MVS holds per-token vectors and scores them, so late interaction is a storage choice rather than an engineering project.
Store multi-vector indexes without building MaxSimQuick Answer
The best overall option in this category is EVIE-Preview-4.5B (Tencent), especially for teams who want the current accuracy ceiling and are storage-constrained enough that token dimension matters more than a preview label.. The rankings below compare each tool by strengths, limitations, pricing, and fit for production use.
EVIE-Preview-4.5B (Tencent)
Best for teams who want the current accuracy ceiling and are storage-constrained enough that token dimension matters more than a preview label..
ColQwen2 and ColQwen2.5 (Vidore)
Best for almost everyone starting out, and anyone who would rather have a well-worn model than the leaderboard leader..
jina-embeddings-v4
Best for pipelines that want a cheap dense first stage and late interaction only on the shortlist..
Skip the comparison? Mixpeek runs visual document retrieval on your own data: extraction, indexing, and search in one platform.
How We Evaluated
Evaluated by the Mixpeek engineering team, who build and operate multimodal retrieval infrastructure in production. Last tested September 2026; rankings re-checked when the market shifts, with pricing and claims verified against each vendor's public documentation.
Retrieval Accuracy
nDCG@10 on ViDoRe V3 public, which is 8 domains crossed with 6 query languages. V1 and V2 are close to saturated at the top of the field, so V3 is the number that still separates these models.
Index Cost Per Million Pages
Vectors per page times token dimension times bytes per value. This is the difference between a 179 GiB index and an 896 GiB one for the same corpus, and it is not on any model card.
Licence
Whether you can ship it. Two strong models on this leaderboard are non-commercial or carry bespoke terms, which removes them from most production shortlists regardless of score.
Ecosystem and Tooling
Whether colpali-engine loads it, whether the vector store supports multi-vector natively, and how much traffic the checkpoint carries, since a model nobody runs has no answers waiting when it breaks.
Quick answer
The short version, before the detail:
- EVIE-Preview-4.5B (Tencent)best for teams who want the current accuracy ceiling and are storage-constrained enough that token dimension matters more than a preview label.Leads the leaderboard at 128-dimension tokens while the model behind it needs 640, which is a 5x index-size difference for 0.04 nDCG.
- ColQwen2 and ColQwen2.5 (Vidore)best for almost everyone starting out, and anyone who would rather have a well-worn model than the leaderboard leader.The only model here with production traffic in the hundreds of thousands per month.
- jina-embeddings-v4best for pipelines that want a cheap dense first stage and late interaction only on the shortlist.Dense and late-interaction from one checkpoint, which is what makes the 24x storage difference a design choice instead of a model choice.
- ColPali v1.3 (Vidore)best for understanding the method, reproducing the paper, or a baseline you can defend in a writeup.The paper everything here descends from, still readable and still runnable.
- webAI ColVec1.1-8bbest for finance and hr document sets specifically, where its per-domain lead is real, and where storage is not the constraint.Ties for the best score in the field, and costs five times the storage to do it.
- NVIDIA nemotron-colembed-vl-8b-v2best for research, internal evaluation, and benchmarking your commercial choice against a strong reference.The strongest model here that you cannot ship.
- ColNomic Embed Multimodal 7Bbest for teams already using nomic atlas who want the visual document index in the same tooling.Apache 2.0 with a mature inspection story around the index rather than just the model.
- Mixpeekbest for teams who have picked a model from this list and do not want to build the multi-vector store, the fusion and the filter layer around it.The layer around the model: multi-vector storage, fusion, reranking and filters, rather than another checkpoint.
Overview
Best Visual Document Retrieval Models: comparison at a glance
| # | Tool | Best for | Pricing | Key differentiator | Main limit |
|---|---|---|---|---|---|
| 1 | EVIE-Preview-4.5B (Tencent) | Teams who want the current accuracy ceiling and are storage-constrained enough that token dimension matters more than a preview label. | Free, Apache 2.0. Self-hosted; you pay for the GPU that indexes and the storage that holds the vectors. | Leads the leaderboard at 128-dimension tokens while the model behind it needs 640, which is a 5x index-size difference for 0.04 nDCG. | Preview, released 2026-08-17, with 2,065 downloads and very little production mileage |
| 2 | ColQwen2 and ColQwen2.5 (Vidore) | Almost everyone starting out, and anyone who would rather have a well-worn model than the leaderboard leader. | Free, Apache 2.0 and MIT. | The only model here with production traffic in the hundreds of thousands per month. | Several points of ViDoRe V3 behind the 2026 models at the top of the leaderboard |
| 3 | jina-embeddings-v4 | Pipelines that want a cheap dense first stage and late interaction only on the shortlist. | Free weights on HuggingFace; Jina also sells it as a hosted embedding API. | Dense and late-interaction from one checkpoint, which is what makes the 24x storage difference a design choice instead of a model choice. | 57.54 on ViDoRe V3 public, nearly 8 points behind the leader |
| 4 | ColPali v1.3 (Vidore) | Understanding the method, reproducing the paper, or a baseline you can defend in a writeup. | Free, MIT. | The paper everything here descends from, still readable and still runnable. | Superseded on accuracy by every 2025 and 2026 model on this list |
| 5 | webAI ColVec1.1-8b | Finance and HR document sets specifically, where its per-domain lead is real, and where storage is not the constraint. | Weights on HuggingFace under an "other" licence. Read the terms. | Ties for the best score in the field, and costs five times the storage to do it. | 640-dimension token vectors, so about 896 GiB per million pages, 5x the 128-dimension models |
| 6 | NVIDIA nemotron-colembed-vl-8b-v2 | Research, internal evaluation, and benchmarking your commercial choice against a strong reference. | Free for non-commercial use only, CC-BY-NC-4.0. | The strongest model here that you cannot ship. | CC-BY-NC-4.0, so not usable in a commercial product |
| 7 | ColNomic Embed Multimodal 7B | Teams already using Nomic Atlas who want the visual document index in the same tooling. | Free, Apache 2.0. Nomic also sells hosted inference. | Apache 2.0 with a mature inspection story around the index rather than just the model. | 57.64 on ViDoRe V3 public, roughly 8 points off the leader |
| 8 | Mixpeek | Teams who have picked a model from this list and do not want to build the multi-vector store, the fusion and the filter layer around it. | Free tier available; usage-based pricing above it. See /pricing. | The layer around the model: multi-vector storage, fusion, reranking and filters, rather than another checkpoint. | No managed extractor runs these models; you bring the vectors or you bring the weights on Enterprise |
EVIE-Preview-4.5B (Tencent)
Ranked first on ViDoRe V3 public at 65.36 nDCG@10 and first on V1+V2 at 85.77, and it does it at 128-dimension token vectors while the model half a point behind it uses 640. That combination is the reason it leads this list rather than the leaderboard: matching the field's best accuracy at a fifth of its index size is a bigger practical difference than the 0.04 points separating the top two. It was trained at 768 visual tokens per page and gains 0.80 nDCG@10 when pushed to 1,792 at test time, with no retraining, improving in 7 of 8 domains. Tencent publishes both tiers: 179.2 GiB per million pages at the training budget, 420.5 GiB at the extrapolated one. Built on Qwen3.5-4B through the ColPali engine, Apache 2.0.
Leads the leaderboard at 128-dimension tokens while the model behind it needs 640, which is a 5x index-size difference for 0.04 nDCG.
Strengths
- +Highest published ViDoRe V3 public score at 65.36, and highest on V1+V2 at 85.77
- +128-dimension token vectors, so a million pages fits in 179 GiB at the training budget
- +Apache 2.0, with the licence text resolving the ambiguity rather than leaving it
- +Loads through colpali-engine, the same path as ColPali and ColQwen
Limitations
- -Preview, released 2026-08-17, with 2,065 downloads and very little production mileage
- -4.54B parameters means a real GPU for indexing, not a laptop
- -The 1,792-token tier that produces the headline score costs 420.5 GiB per million pages
Choose This When
When you are indexing enough pages that the storage arithmetic decides the budget, and you can absorb a preview checkpoint.
Skip This If
When you need a model that has been in production somewhere for a year.
ColQwen2 and ColQwen2.5 (Vidore)
The practical default. ColQwen2-v1.0 carries 342,558 monthly downloads and ColQwen2.5-v0.2 another 216,186, which together are more than the rest of this list combined by two orders of magnitude. Built on Qwen2-VL-2B, it takes dynamic image resolutions without resizing, so aspect ratio survives, with a ceiling of 768 image patches. What you get for the traffic is the thing benchmarks do not show: the failure modes are documented, the integrations exist, and someone has already hit your problem. Apache 2.0 for v1.0, MIT for 2.5.
The only model here with production traffic in the hundreds of thousands per month.
Strengths
- +By far the most-deployed model here, so the tooling and the answers already exist
- +Dynamic resolution without resizing, which preserves aspect ratio on wide tables and scans
- +Apache 2.0 (v1.0) and MIT (v2.5), both unambiguous for commercial use
- +2B parameters, which indexes on far cheaper hardware than the 8B tier
Limitations
- -Several points of ViDoRe V3 behind the 2026 models at the top of the leaderboard
- -128-dimension tokens at up to 768 patches, so the index-cost arithmetic still applies in full
- -Trained on an English-only set, with multilingual ability inherited from the backbone rather than trained for
Choose This When
First. Measure on your own documents, then decide whether the newer models are worth the migration.
Skip This If
When ViDoRe V3 is your acceptance criterion and the gap to 65 matters.
jina-embeddings-v4
The one model here that gives you both modes from one checkpoint. It emits a 2048-dimension single vector, truncatable to 1024, 512, 256 or 128 through Matryoshka, and a 128-dimension multi-vector for late interaction. That matters because the storage difference between the two is roughly 24x, so you can index single-vector, retrieve broadly, and rescore only the shortlist with late interaction. At 467,674 monthly downloads it is the most-used model on this list after ColQwen, and its ViDoRe V3 public score is 57.54.
Dense and late-interaction from one checkpoint, which is what makes the 24x storage difference a design choice instead of a model choice.
Strengths
- +Single-vector and multi-vector from one model, so you can stage cost against accuracy
- +Matryoshka down to 128 dimensions on the dense side, no re-encoding
- +467,674 monthly downloads and task-specific LoRA adapters
- +Handles text, images and visual documents in one space
Limitations
- -57.54 on ViDoRe V3 public, nearly 8 points behind the leader
- -3.75B parameters and no licence field on the HuggingFace record, so check the terms before shipping
- -Two modes means two evaluations, since the dense and late-interaction results differ
Choose This When
When index cost is the binding constraint and you can accept a lower ceiling in exchange for a two-stage design.
Skip This If
When you need the top of the ViDoRe leaderboard.
ColPali v1.3 (Vidore)
The original. ColPali introduced page-image retrieval with late interaction in the paper that named the approach, and every model on this list is a descendant. Built on PaliGemma 3B, MIT licensed, and still the clearest thing to read if you want to understand what these models are doing rather than just run one. 17,227 monthly downloads, which is modest next to ColQwen because the family moved on to Qwen backbones, but the checkpoint is stable and the paper is the reference.
The paper everything here descends from, still readable and still runnable.
Strengths
- +The reference implementation and the paper the whole field cites
- +MIT licence, the most permissive here
- +3B parameters, well-documented behaviour, no surprises
Limitations
- -Superseded on accuracy by every 2025 and 2026 model on this list
- -Fixed resolution, so wide tables and unusual aspect ratios are resized before encoding
- -17,227 monthly downloads against ColQwen's 342,558
Choose This When
As a baseline, or when MIT specifically is the licence you need.
Skip This If
In production, where ColQwen2 is strictly better for the same effort.
webAI ColVec1.1-8b
Second on ViDoRe V3 public at 65.32, four hundredths of a point behind EVIE, and it wins on Finance EN, Finance FR and HR. On accuracy alone it belongs at the top of this list. It is at rank 5 because of what it costs and what you may sign. Its token vectors are 640 dimensions against EVIE's 128, which is 940 KiB per page and roughly 896 GiB per million pages, five times the index for a statistically indistinguishable score. The HuggingFace licence field reads "other", so read the terms before you plan around it. 365 monthly downloads.
Ties for the best score in the field, and costs five times the storage to do it.
Strengths
- +65.32 on ViDoRe V3 public, effectively tied for first
- +Best in the field on Finance EN, Finance FR and HR domains
- +A 4.5B sibling at 63.90 if the 8.4B model is too large
Limitations
- -640-dimension token vectors, so about 896 GiB per million pages, 5x the 128-dimension models
- -Licence reads "other" on the model record, which is not a shipping decision you can make from the card
- -365 monthly downloads and 8 likes, so effectively no production mileage
Choose This When
When your corpus is financial or HR documents and your index budget is elastic.
Skip This If
At scale, or anywhere the licence has to be unambiguous.
NVIDIA nemotron-colembed-vl-8b-v2
63.54 on ViDoRe V3 public, fourth in the field, with a 4B sibling at 61.42 and a 3B Llama-based variant at 59.70. A genuinely strong family, well engineered, with NVIDIA's usual deployment tooling behind it. The licence is CC-BY-NC-4.0. That is non-commercial, which removes it from most production shortlists no matter how it scores, and it is the first thing to check rather than the last. 4,643 monthly downloads and 8.77B parameters.
The strongest model here that you cannot ship.
Strengths
- +63.54 on ViDoRe V3 public, ahead of every 2025 model here
- +Three sizes (8B, 4B, 3B) so you can trade accuracy against indexing cost
- +NVIDIA deployment tooling and NIM packaging
Limitations
- -CC-BY-NC-4.0, so not usable in a commercial product
- -8.77B parameters, the largest model on this list
- -The non-commercial terms mean the ecosystem around it stays thin
Choose This When
Academic work, or as the ceiling you measure your production model against.
Skip This If
Any commercial deployment. The licence is the whole answer.
ColNomic Embed Multimodal 7B
57.64 on ViDoRe V3 public, close to jina-embeddings-v4 and behind the 2026 cohort. Apache 2.0, which puts it in the shippable half of this list, and Nomic ships an ecosystem around it: Atlas for inspecting the embedding space, and a consistent API across their text and multimodal models. 667 monthly downloads, which is low for a model this well known, and worth reading as a signal that most teams evaluating it end up on ColQwen.
Apache 2.0 with a mature inspection story around the index rather than just the model.
Strengths
- +Apache 2.0 at a competitive 2025-era score
- +Nomic's tooling, including Atlas for visualising what the index actually contains
- +Consistent API with their text embedding models, so one client covers both
Limitations
- -57.64 on ViDoRe V3 public, roughly 8 points off the leader
- -667 monthly downloads, far below ColQwen and jina
- -7B parameters for a score two smaller models beat
Choose This When
When you are already in the Nomic ecosystem.
Skip This If
When you are not, since ColQwen2 is smaller, better supported and scores comparably.
Mixpeek is the retrieval layer these models plug into rather than another model. document_graph_extractor@v1 decomposes a PDF into spatial blocks with layout classification, and the retriever runs multi-stage pipelines over whatever vectors a collection holds, so a late-interaction index and a text index answer one query together. Be clear about the boundary, because it decides whether this fits. No managed extractor runs an arbitrary HuggingFace model, and none of the models above is one Mixpeek hosts today. You run the model and upsert the vectors through the documents endpoint, or on Enterprise you upload the weights and load them from a custom plugin. What you do not build is the multi-vector store, the fusion, the reranking stage or the filter layer around them.
The layer around the model: multi-vector storage, fusion, reranking and filters, rather than another checkpoint.
MVS holds the per-token vectors and runs the MaxSim scoring, which is the part of late interaction that a single-vector store cannot do.
Strengths
- +Multi-stage retrievers that fuse a late-interaction index with text, filters and a reranker in one query
- +Multi-vector storage in MVS, so you are not building the MaxSim layer yourself
- +document_graph_extractor gives you the layout blocks alongside the page vectors
- +Object storage is the substrate, so pages stay where they already live
Limitations
- -No managed extractor runs these models; you bring the vectors or you bring the weights on Enterprise
- -The custom plugin path for hosting your own model is Enterprise only
- -A retrieval platform is more moving parts than a single model and a FAISS index
Choose This When
Once the model choice is made and the surrounding system is the work.
Skip This If
When one model and a local index is genuinely all you need.
Which one should you choose?
- Choose EVIE-Preview-4.5B (Tencent) when you are indexing enough pages that the storage arithmetic decides the budget, and you can absorb a preview checkpoint.
- Choose ColQwen2 and ColQwen2.5 (Vidore) if first. Measure on your own documents, then decide whether the newer models are worth the migration.
- Choose jina-embeddings-v4 when index cost is the binding constraint and you can accept a lower ceiling in exchange for a two-stage design.
- Choose ColPali v1.3 (Vidore) if as a baseline, or when MIT specifically is the licence you need.
- Choose webAI ColVec1.1-8b when your corpus is financial or HR documents and your index budget is elastic.
- Choose NVIDIA nemotron-colembed-vl-8b-v2 if academic work, or as the ceiling you measure your production model against.
- Choose ColNomic Embed Multimodal 7B when you are already in the Nomic ecosystem.
- Choose Mixpeek if once the model choice is made and the surrounding system is the work.
Put visual document retrieval to work
Connect a bucket and Mixpeek runs the whole visual document retrieval pipeline for you: extraction, indexing, and search over your own objects. No models to wire up, nothing to host.
Start with ManagedAlready have vectors?
Keep your embeddings on your own cloud and run dense, sparse, and BM25 search directly on object storage. From $25/mo.
Start with MVSFrequently Asked Questions
What does a million pages actually cost to index with late interaction?
Roughly 179 GiB at 128-dimension token vectors and the standard 768-token page budget, before payload. The arithmetic is vectors per page times dimensions times bytes per value: about 750 vectors, 128 dimensions, 2 bytes in bfloat16, which is 188 KiB per page. Tencent publishes 179.2 GiB per million for EVIE at that budget and this reproduces it exactly. A model emitting 640-dimension tokens instead costs about 896 GiB for the same corpus.
How much more does late interaction cost than a single vector?
About 24 times, at the standard budgets. One 2048-dimension float32 vector per page is 8 KiB and 7.6 GiB per million pages. The same page as 750 128-dimension bfloat16 token vectors is 188 KiB and 179 GiB. That is the trade you are making, and it is why jina-embeddings-v4 shipping both modes from one checkpoint is a practical feature rather than a spec-sheet one: index dense, rescore the shortlist with late interaction, pay the multiplier only on the candidates.
Which of these can I actually ship in a commercial product?
EVIE-Preview-4.5B, ColQwen2-v1.0 and ColNomic are Apache 2.0. ColPali v1.3 and ColQwen2.5 are MIT. NVIDIA's nemotron-colembed family is CC-BY-NC-4.0, which is non-commercial, so it is out regardless of its score. webAI ColVec1.1 reads "other" on its model record, which means reading the licence file before you plan around it. jina-embeddings-v4 carries no licence field on the API record at all, so check Jina's terms directly.
Do I still need OCR if I use one of these?
Not for retrieval. That is the point: the model embeds the rendered page, so finding the right page needs no text extraction at all, and it works on scans, charts and handwriting where OCR degrades. You will still want text for anything downstream that reads rather than retrieves, such as quoting an exact figure into an answer or filtering on a field. The common shape is visual retrieval to find the page, then extraction on the handful of pages that came back.
Is ViDoRe V3 the benchmark to use, or V1 and V2?
V3, because V1 and V2 are close to saturated at the top of the field. The leaders on V1+V2 sit around 85, where a point is noise, while V3 spreads the same models from 57 to 65 across 8 domains and 6 query languages. Use V3 to rank and your own documents to decide. The per-domain table matters more than the average: the top two models trade places depending on whether the corpus is Finance, Physics or CompSci.
Can Mixpeek run these models for me?
Not as a managed extractor. GET /v1/discovery/extractors returns 13 extractors and none accepts an arbitrary HuggingFace model id, so there is no field to name one in. Two paths work. Run the model yourself and upsert the per-token vectors through POST /v1/namespaces/{namespace_id}/documents/upsert, where MVS holds them and runs the MaxSim scoring. Or on Enterprise, upload the weights through POST /v1/namespaces/{namespace_id}/models, which accepts the huggingface format, and load them from a custom plugin.
See how Mixpeek handles this
Purpose-built for visual document retrieval models, not bolted on.
Talk to a Mixpeek engineer: free
30 minutes. Bring your use case and we'll tell you exactly what would work and what wouldn't.
Explore Other Curated Lists
Best Rerankers for RAG
A reranker re-scores your first-pass retrieval results so the most relevant ones reach the LLM. We compared the leading 2026 rerankers, managed APIs and open-weight cross-encoders, on relevance lift, latency, license, and language and modality coverage.
Best Reverse Video Search Tools
Reverse video search finds where a clip appears, which videos are near-duplicates, and which library footage is visually similar to a query video. We compared the leading tools on match accuracy, clip and frame-level granularity, index scale, and how they handle re-encodes, crops, and edits.
Best Video Deduplication Tools
How to find duplicate and near-duplicate videos in a library you own. We compare perceptual video hashing, commercial fingerprinting, and embedding similarity on re-encodes, crops, overlays, and clip-length edits.