Tencent Releases EVIE Multilingual Model for Large Language Model Applications

Technical Specifications and Architecture

Tencent has released EVIE-Preview-4.5B, a multilingual Visual Document Retrieval (VDR) model built upon the Qwen3.5-4B architecture. According to technical documentation hosted on Hugging Face, the model utilizes ColBERT-style late interaction with native 128-dimensional multi-vector token embeddings.

The model features 4.54 billion parameters and operates in BF16 precision. Its architecture combines native GatedDeltaNet linear-attention with full-attention hybrid modeling, paired with a compact visual projection. A primary technical objective of the EVIE architecture is the reduction of computational overhead; the model generates ultra-compact 128D multi-vectors, which reportedly reduces vector storage and indexing costs by a factor of 8× to 32× compared to standard 2560D–4096D representations.

Benchmark Performance and Capabilities

Data released via Hugging Face indicates that EVIE-Preview-4.5B achieves high performance across several industry benchmarks. The model reportedly outperforms larger 8B-parameter models on the ViDoRe V3 benchmark, leading in seven out of eight public domains. On the ViDoRe V1+V2 benchmarks, the model achieved an average accuracy of 85.93 nDCG@5.

The model is designed for robust zero-shot generalization across multiple languages, including English, French, German, Italian, Spanish, Portuguese, and Chinese. Its functional scope includes the retrieval of diverse visual formats, such as charts, tables, scientific reports, and financial filings. Furthermore, the model is compatible with the existing colpali-engine ecosystem and late-interaction scoring pipelines.

Practical Applications and Integration

In practical deployment scenarios, the model allows for the embedding of both visual documents (such as PDFs or images) and text queries into a shared vector space. This enables users to search through large repositories of visual data—such as corporate financial filings or scanned images—with low latency.

Developers have noted that when integrated with larger 9B text-generation models, the EVIE architecture can serve as a component in a broader visual document understanding pipeline. While the model is currently available as a preview, its ability to handle complex multi-format documents suggests potential utility in automated data extraction and enterprise-level document management systems.