Tencent EVIE-4.5B: How an OCR-Free Vision Retriever Beat NVIDIA 8B on ViDoRe V3

Forum topic · 小凯 · 2026-08-26

Summary

Tencent's Hunyuan team released EVIE-Preview-4.5B in August 2026, a vision document retrieval model that eliminates OCR entirely and treats each PDF page as a high-fidelity 2D image. Using late interaction with 128-dimensional multi-vector patch embeddings and a GatedDeltaNet linear attention backbone, the 4.5B-parameter model scored 64.40 on ViDoRe V3, surpassing an 8B NVIDIA model that scored 45-52. The article compares the OCR pipeline (image to text to single vector) with the pure-vision pipeline (image directly to compact multi-vector tokens), explains the MaxSim similarity operator, and shows how EVIE acts as a scout in a multimodal RAG system, retrieving the one relevant page in 0.01 seconds before a larger VLM reads it in detail. Reported benefits include roughly 8-32x memory reduction versus 2048-dim embeddings and about 85% storage savings.

Overview

Tencent's Hunyuan team released EVIE-Preview-4.5B in August 2026, a visual document retrieval model that completely removes OCR from the retrieval pipeline. Instead of flattening pages into text, the model encodes each PDF page as a 2D image and uses late interaction with 128-dimensional multi-vector patch embeddings. Despite its small 4.5B-parameter size, it scored 64.40 on ViDoRe V3, outperforming a competing NVIDIA 8B model (45.00–52.00) on the same benchmark.

Key Points

1. OCR vs. Pure-Vision Pipelines

OCR pipeline: Image → OCR text → flat embedding → search. Tables break across rows, charts are lost, layout is destroyed.
EVIE-4.5B pipeline: Image → direct visual encoding → 128-dim compact multi-vector tokens per patch → millisecond retrieval.

2. Core Mechanisms

Late Interaction (ColPali paradigm): Each page patch keeps its own embedding instead of being averaged into one global vector. Query tokens are matched against patches at search time.
MaxSim similarity operator:

\text{Score}(q, d) = \sum_{i \in \text{query}} \max_{j \in \text{patches}} \left( E_q[i] \cdot E_d[j]^T \right)

128-dim compression: Patches are encoded into 128 dimensions instead of the typical 2048, reducing memory by 8–32x and storage cost by ~85%.
GatedDeltaNet linear attention hybrid: A more efficient backbone than full Transformer attention, enabling high-throughput processing of long documents.

3. Multimodal RAG Role

4. Reported Results

ViDoRe V3 score: 64.40 (1st place among compared models).
Beats: NVIDIA's 8B model (45.00–52.00).
Storage: ~85% reduction vs. 2048-dim embeddings.
Speed: ~0.01 s per query across 100k pages.

Academic References

1. Khattab & Zaharia, SIGIR 2020 — *ColBERT: Efficient and effective passage search via contextualized late interaction over BERT*. Founded late interaction and MaxSim theory. 2. Faysse et al., 2024 — *ColPali: Efficient Document Retrieval with Vision Language Models*. DOI/arXiv: arXiv:2407.01449. Established the ViDoRe benchmark and end-to-end vision-based document retrieval paradigm.