> ## Content Index
> Fetch the complete content index at: https://globalfeed.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# Tencent quietly releases EVIE, a model that searches documents visually
- URL: https://globalfeed.ai/en/tencent-quietly-releases-evie-a-model-that-searches-documents-visually/
- Published: 2026-08-28T00:49:05.000Z
- Updated: 2026-08-28T13:42:08.000Z
- Description: The 8.5 GB open-source model searches the page itself rather than converting PDFs, tables and scanned files to text. It outranks models twice its size on ViDoRe and runs locally.
- Author: GlobalFeed Editor
- Tags: Tencent, open source, document retrieval, AI, Qwen, dil-en, elle

Tencent uploaded EVIE-Preview-4.5B, a visual document retrieval model, to Hugging Face without an announcement. It is released under Apache 2.0, so the weights can be downloaded and run locally.

The model is not a chat assistant. Its job is retrieval: finding the page in a pile of documents that answers a question. What sets it apart is that it does not convert the page to text first. It looks at the page itself, with its charts, tables, layout and images intact. Tencent says it retrieves over charts, tables, scientific reports, financial filings and scanned forms, across multiple languages.

Technically the model is built on the Qwen3.5-4B vision-language backbone and carries 4.54 billion parameters; the BF16 checkpoint is 8.5 GB. Instead of text chunks it produces 128-dimensional multi-vectors per page.

According to figures the company published, the model ranks first on ViDoRe V3 with 65.36 nDCG@10, against 65.32 for second place. On ViDoRe V1+V2 it leads across 14 tasks with 85.77 nDCG@5, against 84.90\. These numbers come from the model card itself and are not independently verified.

Keeping the vector dimension small matters for storage. In Tencent's own example, a raw index of one million pages takes 179.2 GiB at the 768-token-per-page budget, and 420.5 GiB at the 1,792-token tier.

For the model card and weights see the [Hugging Face page](https://huggingface.co/tencent/EVIE-Preview-4.5B?ref=globalfeed.ai), and for the code, the [GitHub repository](https://github.com/Tencent/EVIE-Preview-4.5B?ref=globalfeed.ai).