EVIE-8B
by tencent
The largest EVIE visual document retriever, and the teacher behind EVIE-4.5B
tencent/EVIE-8BDeploy EVIE-8B
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-8B finds the page in a pile of PDFs, slides and scans that answers a question, working from an image of each page with no OCR step. It is a late-interaction retriever: every page becomes a set of 4096-dimension vectors, one per visual token, and a query is scored against a page by matching each query token to its closest page vector (MaxSim). Tencent released it on 4 September 2026 under Apache-2.0, with 8.4B parameters on a Qwen3.5 backbone.
The model card reports 66.75 nDCG@10 on ViDoRe V3 under Tencent's own evaluation protocol, and lists it first on the ViDoRe V3 leaderboard at a mean task score of 66.24. Its smaller sibling, EVIE-4.5B, reports 66.02 on the same protocol, and was distilled from this model.
The cost is storage. Each page vector is 4096 floats, and in the card's own example one page produced 3,161 of them, about 49 MiB per page at float32. For most indexes EVIE-4.5B, whose vectors shrink to 64 dimensions, is the practical choice, and EVIE-8B is the reference for how much accuracy that gives up.
Architecture
A ColQwen3.5 model: the Qwen3.5 vision encoder turns a page image into patch tokens, the language model runs over them with bidirectional attention enabled, and a projection emits one 4096-dimension vector per token. Queries go through the same model as text. Relevance is the sum, over query tokens, of each token's maximum dot product with any page vector. The default page budget allows up to 16,384 visual tokens; the card's evaluation protocol uses 1,024. Training used 775,635 document-query pairs with hard negatives sorted into answerable, ambiguous and strict, and the released checkpoint is a weight-space mix (alpha 0.40) of two separately trained runs.
Mixpeek SDK Integration
// EVIE-8B returns a SET of 4096-dimension vectors per page, scored with MaxSim.
// No Mixpeek extractor runs these weights, and the documents upsert endpoint 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. MVS scores multi-vector
// indexes with MaxSim.
//
// Pages enter as images in a bucket whose schema has one image property.
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
- Page retrieval from images, with no OCR step
- Late-interaction scoring (MaxSim) over 4096-dimension token vectors
- Loads through colpali-engine or Sentence Transformers (MultiVectorEncoder)
- Apache-2.0 license, with training code and the evaluation harness in Tencent/EVIE
Use Cases on Mixpeek
Benchmarks
| Dataset | Metric | Score | Source |
|---|---|---|---|
| ViDoRe V3 | nDCG@10 | 66.75 | Model card: tencent/EVIE-8B (self-reported, Tencent's protocol dated 2026-08-27) |
| ViDoRe V3 leaderboard | Mean (task) | 66.24 | Model card, citing the ViDoRe leaderboard |
| ViDoRe V2 | nDCG@5 | 74.23 | Model card (self-reported) |
| ViDoRe V1 | nDCG@5 | 92.18 | Model card (self-reported) |
| JinaVDR (76 tasks) | nDCG@10 | 83.30 | Model card (self-reported) |
Performance
Index size is the constraint. In the card's Sentence Transformers example a page produced 3,161 vectors, which at 4096 float32 dimensions is about 49 MiB per page; the evaluation protocol's 1,024-token budget caps a page at far fewer. We have not measured encoding latency; the card loads the model in bfloat16 with flash_attention_2.
Common Pipeline Companions
Frequently Asked Questions
Should I use EVIE-8B or EVIE-4.5B?
For most indexes, EVIE-4.5B. On the card's own protocol it scores 66.02 nDCG@10 on ViDoRe V3 against 66.75 for EVIE-8B, and its vectors can be cut to 64 dimensions and compressed to 32 per page. EVIE-8B keeps 4096-dimension vectors for every token, which is the reason to use it when storage is not the constraint, or as the reference when you measure what a smaller retriever loses on your documents.
How much storage does EVIE-8B need per page?
It depends on how many visual tokens a page produces. Each vector is 4096 floats, 16 KiB at float32. In the card's own example one page produced 3,161 vectors, about 49 MiB. The evaluation protocol caps pages at 1,024 visual tokens, which bounds that at 16 MiB, still large enough that most teams index with a smaller model.
Are EVIE-8B's benchmark numbers independently verified?
The 66.75 nDCG@10 figure is self-reported under Tencent's own protocol. The card separately cites the ViDoRe leaderboard, where it is listed first at a mean task score of 66.24. The benchmarks are public, so the numbers can be reproduced; until you have, test it on a sample of your own documents.
Does Mixpeek run EVIE-8B?
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 MVS scores multi-vector indexes with MaxSim.
Specification
Research Paper
Tencent/EVIE on GitHub (the card says the formal paper is to follow)
arxiv.orgBuild a pipeline with EVIE-8B
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