EVIE-4.5B
by tencent
Search document pages as images, charts and tables included, with a per-page vector count you can shrink to 32
tencent/EVIE-4.5BDeploy EVIE-4.5B
Single-tenantMixpeek has no managed extractor for this model. On a single-tenant deployment you upload the weights and a custom plugin serves them next to the rest of your pipeline.
Overview
EVIE-4.5B searches document pages as images. It turns each page into a set of vectors from its visual patches and scores a query by matching every query token against its best page vector and summing those matches, a method called MaxSim. Charts, table layouts and small type stay in that representation, which matters for scanned reports and slide decks whose text layer is missing or wrong. Tencent released it on 4 September 2026, and its card reports 66.02 nDCG@10 on ViDoRe V3.
The cost of this approach is storage, and the release addresses it twice. The output projection is 2048-dimensional and can be truncated at query time to 1024, 512, 256, 128 or 64 dimensions; the card's table shows ViDoRe V3 moving from 66.02 at 2048 to 64.51 at 64. A training-free clustering step called HAC also reduces the roughly 750 vectors per page to 32 or 64. In the smallest configuration the card reports 3.81 GiB per million pages, with ViDoRe V3 at 59.58.
Weigh the evidence before adopting it. Every benchmark figure is self-reported on the card, the formal paper is announced for a later release, and the repository showed 1,035 downloads in the month before this page was written. Test it on a sample of your own pages and treat the reported numbers as the vendor's claim.
Architecture
ColQwen3_5: a Qwen3.5 language backbone (hidden size 2560, 32 layers, 248,320-token vocabulary) with a vision encoder (hidden size 1024, depth 24, projected to 2560), run with bidirectional attention and a 2048-dimensional multi-vector projection head. The safetensors index totals 4,544,510,464 BF16 parameters; the card's comparison table lists 4.61B. The sentence-transformers config declares a MultiVectorEncoder with maxsim similarity. The card describes it as distilled from EVIE-8B, with tencent/EVIE-Preview-4.5B as the base checkpoint.
Mixpeek SDK Integration
// EVIE returns a SET of vectors per page, and the score is MaxSim across them.
// No Mixpeek extractor runs these weights, and the documents upsert endpoint in
// the public API reference takes one vector per named index. The path that keeps
// the multi-vector form is a single-tenant Enterprise deployment: upload the
// weights through POST /v1/namespaces/{id}/models, and a custom plugin declares
// feature_type multi_vector and returns one list of vectors per page
// (list[list[float]]). MVS scores multi-vector indexes with MaxSim.
//
// The pages enter the way any object does: as images in a bucket whose schema
// has one image property, here named "page".
const res = await fetch(
"https://api.mixpeek.com/v1/buckets/bkt_your_bucket/objects",
{
method: "POST",
headers: {
Authorization: "Bearer API_KEY",
"X-Namespace": "ns_your_namespace",
"Content-Type": "application/json",
},
body: JSON.stringify({
blobs: [{ property: "page", type: "image", data: pageImageUrl }],
}),
},
);Capabilities
- Late-interaction retrieval over page images, scored with MaxSim
- One 2048-dimension head truncatable to 1024, 512, 256, 128 or 64
- Training-free HAC compression from roughly 750 to 32 or 64 vectors per page
- Apache-2.0 license, with weights, training and compression code published
Use Cases on Mixpeek
Benchmarks
| Dataset | Metric | Score | Source |
|---|---|---|---|
| ViDoRe V3 | nDCG@10 | 66.02 | Model card: tencent/EVIE-4.5B (self-reported) |
| ViDoRe V1 | nDCG@5 | 92.07 | Model card: tencent/EVIE-4.5B (self-reported) |
| ViDoRe V2 | nDCG@5 | 73.38 | Model card: tencent/EVIE-4.5B (self-reported) |
Performance
Index size is the number to plan for. Per the card, an uncompressed page carries roughly 750 vectors, each 4 KiB at 2048 dimensions, and the d64 K32 configuration stores 32 vectors per page at 3.81 GiB per million pages. We have not measured encoding latency; the card's example loads the model with flash_attention_2.
Common Pipeline Companions
Frequently Asked Questions
What is late interaction, and why use it for document pages?
A single-vector model compresses a whole page into one point. A late-interaction model keeps one vector per region of the page and, at query time, matches each query token against its closest page vector and adds up those best matches. Small details such as one cell in a table or a label on a chart keep their own vectors, so a query about them has something specific to match.
How much storage does EVIE-4.5B need per page?
It depends on the configuration. At the full 2048 dimensions each vector is 4 KiB and an uncompressed page has roughly 750 of them, which is about 2.9 MiB per page. With 64-dimension vectors and HAC compression to 32 vectors per page, the card reports 4 KiB per page and 3.81 GiB per million pages, at a ViDoRe V3 score of 59.58 against 66.02 for the full representation.
Are EVIE's benchmark numbers independently verified?
They are self-reported on the model card, measured with an evaluation protocol the card names and dates 2026-08-27, and the card says the formal paper will follow in a later release. The ViDoRe benchmarks are public, so the scores can be reproduced. Until they are, run it on a sample of your own documents.
Does Mixpeek run EVIE-4.5B?
Not on the managed tier. On a single-tenant Enterprise deployment the weights can be uploaded and run by a custom plugin that declares feature_type multi_vector and returns a list of vectors per page, and MVS scores multi-vector indexes with MaxSim.
Specification
Build a pipeline with EVIE-4.5B
Add this model to a processing pipeline alongside other extractors. Combine with retrieval stages for end-to-end search.
Run it on your own data, free