AI Daily · 2026-09-07

Tencent Hunyuan released two open-source visual document retrieval models, EVIE-8B and EVIE-4.5B, both built on Qwen3.5 and licensed under Apache-2.0 with weights, training, and evaluation code published. The models use late-interaction multi-vector representations to retain fine-grained layout, table, and figure information, scoring 66.75 and 66.02 nDCG@10 on ViDoRe V3 and ranking ahead of comparable models across 138 multilingual tasks. The 4.5B variant is distilled from the 8B teacher via capability-aware relation and margin objectives, and supports Prefix-MRL truncation plus HAC token compression down to about 32 tokens per page, reducing a million-page index to 3.81 GiB. Together, they cover both high-accuracy retrieval and lightweight, storage-efficient deployment.

East Asia · First-hand

Tencent Hunyuan

⭐⭐⭐ [Model Release] tencent/EVIE-8B

Tencent Hunyuan released EVIE-8B, an 8B visual document retrieval model based on Qwen3.5, fully open-sourced under Apache-2.0 including model weights, inference pipelines, evaluation suites, training code, and distillation recipes. It employs bidirectional attention and 4096-dimensional per-token multi-vector embeddings, using late-interaction MaxSim matching to preserve fine-grained layout, typography, charts, and table structure. EVIE-8B achieves 66.75 nDCG@10 on ViDoRe V3, and is validated across 138 tasks covering ViDoRe V1/V2/V3 and JinaVDR, outperforming other multi-vector visual document retrievers. It also acts as the flagship teacher model, providing capacity-aware relation and margin distillation targets to train the lighter EVIE-4.5B Prefix-MRL model.

⭐⭐⭐ [Model Release] tencent/EVIE-4.5B

Tencent Hunyuan released open-source visual document retrieval models EVIE-4.5B and EVIE-8B (based on Qwen3.5). Using a late-interaction multi-vector paradigm, they feature single-projection Prefix-MRL supporting runtime truncation from 2048D down to 64D, and training-free HAC token compression that cuts tokens per page from about 750 to 32, reducing 1M-page index storage to 3.81 GiB. On ViDoRe V3, the 8B model scores 66.75 and the 4.5B scores 66.02; models were evaluated across 138 multilingual tasks on ViDoRe V1/V2/V3 and JinaVDR. The 4.5B student is distilled from the 8B teacher using ARD. All weights, training pipelines, compression algorithms, and evaluation suites are open-sourced under Apache-2.0.