Tencent open-sourced EVIE-Preview-4.5B, a visual document retrieval model
Tencent released EVIE-Preview-4.5B, a 4.54B-parameter visual document retrieval model, on Hugging Face under Apache-2.0. It ranks #1 on ViDoRe V3 with 65.36 nDCG@10 and #1 on ViDoRe V1+V2 with 85.77 nDCG@5. The model uses native 128-dimensional token vectors and is built on Qwen3.5-4B with the ColPali engine.
China context
- Original name
- 腾讯
- Outside China
- Open weights · huggingface.co
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Builders can integrate a state-of-the-art visual document retriever with Apache-2.0 license and 128D vectors, reducing index storage costs.
- For investors
- Tencent's open-weights release in visual document retrieval may pressure commercial document AI vendors and signal a shift toward open multimodal retrieval models.
Tencent published EVIE-Preview-4.5B on Hugging Face, a visual document retrieval model with 4.54B parameters, Apache-2.0 license, and ColPali engine. The model card reports rank #1 on ViDoRe V3 (65.36 nDCG@10) and ViDoRe V1+V2 (85.77 nDCG@5), with native 128-dimensional token vectors. It supports 7 query languages and offers two deployment tiers: 768 visual tokens per page (64.56 V3 public, 179.2 GiB per 1M pages) and 1,792 tokens (65.36 V3 public, 420.5 GiB per 1M pages). The model is based on Qwen3.5-4B. The page also notes newer official models EVIE-4.5B and EVIE-8B, but this evidence focuses on the preview checkpoint.
EVIE-Preview-4.5B uses native 128-dimensional token vectors, reducing index size compared to higher-dimensional alternatives. The model was trained at 768 visual tokens per page; the 1,792-token tier is test-time extrapolation with the same weights, improving nDCG@10 by 0.80 over the training budget. Index cost is 420.5 GiB per 1M pages at BF16 for the extrapolated tier.
Developers outside China can now use a top-ranked visual document retriever under Apache-2.0, potentially lowering the cost of building document search systems compared to proprietary APIs. Competitors like webAI and NVIDIA face a new open-weights benchmark leader on ViDoRe V3.
Open-weights Apache-2.0 license allows commercial use without per-seat fees, and the 128D vectors reduce storage costs for large document collections. The model's multilingual support (7 query languages) may broaden its applicability in international deployments.
The model card points to newer official models EVIE-4.5B and EVIE-8B; a specific signal to check is whether those models are released on Hugging Face and whether their ViDoRe scores are independently reproduced.