tencent / EVIE-Preview-4.5B

---

Overview

EVIE-Preview-4.5B is a state-of-the-art multilingual Visual Document Retrieval (VDR) model built upon Qwen3.5-4B . It employs ColBERT-style late interaction with native 128-dimensional multi-vector token embeddings (4.54B parameters, BF16).

By combining native GatedDeltaNet linear-attention and full-attention hybrid modeling with a compact visual projection, EVIE achieves top-tier performance across ViDoRe V1+V2 and ViDoRe V3 while generating compact 128D multi-vector representations.

Note : This is a preview release. The next iteration of EVIE is pending release.

Key Highlights

• 🎯 Ultra-Compact 128D Token Vectors : Native 128-dimensional multi-vector representations keep each token vector narrow while preserving strong retrieval quality.
• 🏆 SOTA on ViDoRe Benchmarks : Outperforms larger 8B models on ViDoRe V3 (leading 7 of 8 public domains) and delivers top average accuracy on ViDoRe V1+V2 ( 85.93 nDCG@5).
• 🌐 Robust Multilingual & Multi-Format : Strong zero-shot generalization across diverse languages (EN, FR, DE, IT, ES, PT, ZH, etc.) and visual formats (charts, tables, scientific reports, financial filings).
• ⚡ Seamless ColPali Compatibility : Fully integrated with the standard colpali-engine ecosystem and late-interaction scoring pipelines.

---

Architecture

  Text Query ───────► ColQwen3_5 (BiDir Attn) ─────► Query Token Embeddings (128D)
                                                                 │
                                                       Late Interaction (MaxSim) ──► Relevance Score
                                                                 │
Document Image ─────► ColQwen3_5 (Dynamic Vision) ──► Doc Token Embeddings (128D)

1. Vision-Language Backbone : Built on Qwen3.5-4B with interleaved linear and full attention layers.
2. Compact Projection : Projects contextual token states directly into 128-dimensional representations.
3. Late-Interaction Retrieval : Calculates similarity via token-level MaxSim operator across query tokens and document visual tokens.

---

Benchmark Results

Model Footprint

Relative width compares the native uncompressed width of one token vector with EVIE's 128D output. It is not a total index-size comparison: total storage also depends on vectors per page, data precision, projection or compression settings, and index overhead.

---

ViDoRe V3: 8 Public Domains (nDCG@10)

Evaluated across 8 domains with queries spanning 6 languages (EN, FR, DE, IT, PT, ES):

🏆 Result : EVIE leads in 7 out of 8 public domains on ViDoRe V3.

---

ViDoRe V1 + V2 (nDCG@5)

Avg : Unweighted mean across all 14 tasks. Tasks 1–10: ViDoRe V1. Tasks 11–14: ViDoRe V2.

---

Quick Start

Installation

pip install -r requirements.txt

Python Inference

import torch
from PIL import Image
from colpali_engine.models import ColQwen3_5, ColQwen3_5Processor

# Use the Hugging Face model repository or local directory
model_id = "tencent/EVIE-Preview-4.5B"

# 1. Load model and enable bidirectional attention
model = ColQwen3_5.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="cuda",
    attn_implementation="flash_attention_2",
).eval()
model.enable_bidirectional_attention()

# 2. Load processor
processor = ColQwen3_5Processor.from_pretrained(model_id)

# 3. Prepare inputs
images = [Image.open("document_page.png")]
queries = ["What key insights are presented on this page?"]

image_batch = processor.process_images(images).to(model.device)
query_batch = processor.process_queries(queries).to(model.device)

# 4. Generate multi-vector embeddings and score
with torch.inference_mode():
    image_embeddings = model(**image_batch)
    model.rope_deltas = None  # required before query forward
    query_embeddings = model(**query_batch)

scores = processor.score(query_embeddings, image_embeddings)
print("Late-interaction retrieval scores:", scores)

⚠️ Important : Both model.enable_bidirectional_attention() and resetting model.rope_deltas = None prior to query forward passes are required to replicate the checkpoint's full retrieval performance.

CLI Scoring Tool

You can also use the included infer.py script directly:

python infer.py --query "Quarterly revenue report" --image document_page_1.png --image document_page_2.png

---

Reproducing

Run the end-to-end evaluation benchmark across all visible GPUs:

bash reproduce.sh

Notes

• Automatic Dataset Download : On the first run, reproduce.sh automatically invokes download_data.py to fetch the 22 public ViDoRe datasets (~55 GB) from Hugging Face.
• Custom Dataset Path : To reuse an existing dataset directory, pass it directly:

  bash reproduce.sh /path/to/vidore
  

• Target Aggregates :

ViDoRe V1        nDCG@5    91.72  (10 tasks)
ViDoRe V2        nDCG@5    71.44  (4 tasks)
ViDoRe V1+V2     nDCG@5    85.93  (14 tasks)
ViDoRe V3 public nDCG@10   64.40  (8 domains x 6 languages)

---

Training Details

EVIE was trained on approximately 0.8 million high-quality image-query pairs covering multilingual documents, technical reports, complex financial tables, infographics, and document visual QA.

Hard Negative Mining & Data Filtering

• Dynamic Mining & Verification : Hard negatives are actively mined using intermediate retrievers and re-verified:
  • Candidates that accurately answer the query are promoted to positives .
  • Partially relevant or ambiguous candidates are masked out of the loss.
  • Only strictly irrelevant pages are retained as true hard negatives .
• Quality Filtering : Rows containing empty queries, corrupted images, or degraded text are systematically discarded.

---

Acknowledgements

• Built upon the ColPali Engine developed by Illuin Technology.
• Powered by the Qwen3.5-4B vision-language backbone.
• Evaluated on the ViDoRe Benchmark family.

---

Citation

@misc{tencent2026evie,
  title        = {EVIE-Preview-4.5B},
  author       = {{Tencent}},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/tencent/EVIE-Preview-4.5B}},
  note         = {Multilingual visual document retrieval with compact multi-vector embeddings}
}