Tencent open-sourced EVIE-4.5B visual document retriever
Tencent released EVIE-4.5B, a 4.5B-parameter visual document retrieval model, on Hugging Face under Apache 2.0. It achieves 66.02 nDCG@10 on ViDoRe V3 and compresses index storage to 3.81 GiB per million pages.
China context
- Original name
- 腾讯
- Outside China
- Open weights · huggingface.co
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers can integrate EVIE-4.5B into retrieval pipelines with flexible embedding dimensions and low storage overhead, using the Apache 2.0 license for commercial applications.
- For investors
- Tencent's release of a top-performing open-weights model in visual retrieval may pressure competitors and attract enterprise adoption, potentially shifting market share in document AI.
Tencent released EVIE-4.5B, a 4.5B-parameter visual document retrieval model, on Hugging Face under Apache 2.0. The model uses a late-interaction multi-vector paradigm with Prefix-MRL for elastic embedding dimensions (64–2048D) and training-free HAC token compression, reducing index storage to 3.81 GiB per million pages. It achieves 66.02 nDCG@10 on ViDoRe V3, ranking second on the leaderboard behind Tencent's own EVIE-8B.
EVIE-4.5B employs a single 2048D linear projection that supports runtime truncation to {64, 128, 256, 512, 1024, 2048} dimensions without separate checkpoints. HAC (Hierarchical Agglomerative Clustering) compresses visual patch tokens from ~750 to 32 vectors per page, drastically reducing index size. The model is distilled from an 8B teacher using Anchor-preserving Relation Distillation (ARD).
Developers outside China can now use a state-of-the-art visual document retriever with a permissive Apache 2.0 license, reducing infrastructure costs for large-scale document indexing. Competitors in visual retrieval must match or exceed EVIE-4.5B's accuracy and storage efficiency to remain competitive.
EVIE-4.5B enables cost-effective visual document retrieval at scale, with a 3.81 GiB per million pages index footprint. Its open-source license allows commercial use without licensing fees, potentially lowering total cost of ownership for enterprises.
The next verifiable signal is whether Tencent releases the formal research paper and training pipelines mentioned in the model card. Additionally, adoption can be tracked via downloads and community forks on Hugging Face.