Tencent EVIE-8B

EVIE-8B, published on 4 September 2026, does not answer questions. It is a retriever: it converts a scanned page, a slide or a chart into numbers so that a search engine can find the right document. What separates it from an ordinary embedding model is that it never converts the page into text. The page is processed as an image, and a query is matched against the picture itself. The reason matters for anyone who has tried to search a folder of PDFs. Conventional retrieval first runs a document through OCR, then compresses the extracted text into a single vector. Both steps throw information away: OCR loses the layout, and the single vector loses everything that did not fit into one point in space. EVIE keeps a separate vector for every token of the page — roughly 750 of them per page, 4,096 numbers each — and scores relevance by matching each query token against the best-fitting part of the page. Tencent calls this late interaction; the technique comes from the ColBERT and ColPali line of work, and EVIE is a ColQwen3.5 model, built on Alibaba's Qwen3.5-9B backbone rather than on anything of Tencent's own. On the vendor's own comparison table it is the most accurate retriever listed: 66.75 nDCG@10 on ViDoRe V3, 92.18 on V1 and 74.23 on V2, ahead of webAI, Nemotron and Tomoro models of the same size. Tencent evaluated it across 138 tasks drawn from ViDoRe V1, V2, V3 and JinaVDR, which is a wider bench than most competitors publish. The cost is storage, and the reader should weigh it before choosing this model over its smaller sibling. Keeping 750 vectors of 4,096 numbers per page is an expensive way to hold an archive, and Tencent publishes no index-size figure for the 8B at all — the 3.81 GiB per million pages quoted in the family's materials belongs to EVIE-4.5B after token compression. In practice Tencent positions the 8B as a teacher: the model from which the deployable 4.5B student is distilled, and which loses only 0.73 points of accuracy in the process. Weights are Apache 2.0 with no regional carve-out. The technical paper had not been published at release — the model card promises architectural details and ablations in a later update, which means the numbers above rest on the vendor's word alone. Adoption in the 30 days to 16 September 2026 is modest: 596 Hugging Face downloads and 17 likes, well below the 2,614 downloads of the family's August preview.

▸Videos

No videos yet.

▸Reviews