Pure-vision retrieval tops ViDoRe V3 leaderboard: GLM agent + EVIE hits 70.01 nDCG@10
CShorten30 · x · 2026-11-20
A Hugging Face community article presents two retrieval pipelines that work on page images only, both running open-weight models on a single GPU, topping the ViDoRe V3 pipeline leaderboard:
- VisAgentRet: GLM-5.3-Flash uses EVIE-4.5B as a search tool, issuing its own queries and reading candidate pages — 70.01 average nDCG@10, the highest score at publish time.
- EVIE-4.5B + JEV-27B-VL reranker: reranks top-16 page images in one forward pass — 68.25 at 1.47s per query, ranked third.
Both share a first stage built on Tencent's EVIE-4.5B, a late-interaction retriever encoding each page as 1,000 visual token vectors. The result shows pure-vision pipelines can match or beat the best text-only ones. Code for both is Apache-2.0 on GitHub.
More from Multimodal
- Turning Sonnet 3's glossolalia into hyperpop songs with Suno — repligate · 2026-11-20
- Dev builds pixel arcade shooter with Astra 6.2 and Scenario in a few rounds — AIandDesign · 2026-11-20
- Synthesia lands in Claude: agentic video generation inside chat — synthesiaIO · 2026-11-20
- AI cover of 'Kiss from a Rose' impresses past the first seconds of jank — rickasaurus · 2026-11-20
- Krea Bots beta draws 50,000+ applications; new testers get $100 in credits — angrypenguinPNG · 2026-11-20
- Local head-to-head: H3 Image Edit vs Qwen Image 2.1 Edit on an RTX 5090 — Producing_It · 2026-11-19