Tasks and inputs
EVIE 4.5B is listed for Embeddings, Visual document retrieval. It accepts text, image and produces embedding. Use the publisher’s input format and preprocessing for your chosen deployment.
SyncDev model directory · Profile updated Oct 7, 2026
EVIE 4.5B is a visual-document retrieval model that scores page relevance using multiple vectors and late interaction.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
EVIE 4.5B is a visual-document retrieval model that scores page relevance using multiple vectors and late interaction.
EVIE encodes queries and page images as multiple vectors and uses late-interaction scoring, rather than producing one document vector for a conventional cosine-similarity index. The 4.5B Apache-licensed checkpoint uses 2048-dimensional token vectors. The card documents Matryoshka prefix dimensions and hierarchical token compression. Test those options against your labelled queries before reducing stored vectors.
Build a collection with tables, figures and text-rich pages, then label the pages needed to answer realistic questions. Include queries that refer to a chart relationship or a value’s row and column. Assess retrieval before asking a language model to compose an answer. A relevant-looking page can still be the wrong source for a particular quantity.
Use a retrieval engine that supports the documented multi-vector representation and MaxSim-style scoring. A single-vector database path is not automatically a drop-in replacement. Keep query encoding, page preprocessing, dimensions and compression settings aligned across the index and application. Measure retrieval latency and storage at the expected collection size.
The model returns retrieval representations and scores, not prose. If an answer-generation stage follows retrieval, keep references to the actual selected pages and evaluate that stage separately. Changes in scoring backend or numeric precision should be checked against the same labelled query set.
Compare the model on the work you need to complete.
EVIE 4.5B is listed for Embeddings, Visual document retrieval. It accepts text, image and produces embedding. Use the publisher’s input format and preprocessing for your chosen deployment.
Evaluate retrieval over a labelled query set. Keep the encoder, vector dimensions and index settings consistent.
The weights are downloadable under Apache 2.0. Check runtime support and size the complete model and working context for your expected concurrency.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
athrael-soju
athrael-soju/colqwen3.5-4.5B-v3 is a source-linked Hugging Face repository assigned to Visual document retrieval. Its recorded task interface accepts document, text and produces ranked documents.
Tencent
EVIE 8B is a visual-document retrieval model that scores page relevance using multiple vectors and late interaction.
Haon-Chen
Haon-Chen/e5-omni-7B is a source-linked Hugging Face repository assigned to Visual document retrieval. Its recorded task interface accepts document, text and produces ranked documents.
jinaai
jinaai/jina-embeddings-v4 is a source-linked Hugging Face repository assigned to Visual document retrieval. Its recorded task interface accepts document, text and produces ranked documents.
jinaai
jinaai/jina-embeddings-v4-vllm-retrieval is a source-linked Hugging Face repository assigned to Visual document retrieval. Its recorded task interface accepts document, text and produces ranked documents.
lightonai
lightonai/MonoQwen2-VL-v0.1 is a source-linked Hugging Face repository assigned to Visual document retrieval. Its recorded task interface accepts document, text and produces ranked documents.