Tencent open-sourced EVIE-8B visual document retriever
Tencent released EVIE-8B, an 8.4B-parameter visual document retrieval model, on Hugging Face under Apache 2.0. It achieves 66.75 nDCG@10 on ViDoRe V3, ranking first on the leaderboard. Weights, inference pipelines, and evaluation suites are open-sourced.
China context
- Original name
- 腾讯
- Outside China
- Open weights · huggingface.co
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Builders outside China can download the Apache 2.0 weights and integrate EVIE-8B into visual document retrieval pipelines without licensing fees.
- For investors
- Investors should note that Tencent is releasing competitive open-weights models, which may pressure commercial retrieval API providers and shift enterprise spending toward self-hosted solutions.
EVIE-8B uses late-interaction token embeddings with 4096-dimensional per-token representations and full bidirectional attention across multimodal vision-text sequences. It serves as a teacher model for the smaller EVIE-4.5B via topological relation transfer and hard-negative margin supervision.
Developers outside China can now use a state-of-the-art visual document retriever under Apache 2.0, reducing the cost and effort of building high-accuracy document search systems. Competitors must match its 66.75 nDCG@10 on ViDoRe V3 or risk losing users to a free, open-weights alternative.
EVIE-8B provides a high-accuracy, open-weights visual document retrieval model that can be integrated into enterprise search, document management, and RAG pipelines without licensing fees. Its top leaderboard position may attract developers and companies seeking to improve retrieval quality.
The model card states that full technical details, architectural ablations, and the formal research paper will be updated in an upcoming release. A specific signal to watch is whether Tencent publishes the paper and whether the EVIE-4.5B student model gains adoption in production systems.