Tencent EVIE-4.5B

EVIE-4.5B, published on 4 September 2026 alongside the larger EVIE-8B, searches documents by looking at them. It does not run text recognition first: a scanned page, a slide or a table is embedded as an image, and a written query is matched against the picture. The technique — a separate vector per token, scored by best-match against the query, known as late interaction — is accurate but notoriously expensive to store, and this model is Tencent's answer to that expense. Two mechanisms do the work. The first, which Tencent calls Prefix-MRL, trains a single 2,048-dimension projection in such a way that it can simply be cut short at run time: an operator picks 64, 128, 256, 512, 1,024 or 2,048 numbers per vector without swapping models or retraining. The published figures show how gentle the trade is — at the full 2,048 dimensions the model scores 66.02 on ViDoRe V3, at 256 dimensions 65.68, and at 64 dimensions 64.51, while on the older ViDoRe V1 benchmark the 64-dimension setting actually reads 92.16 against 92.07 for the full projection, a difference inside the noise. The second mechanism attacks the number of vectors rather than their length. Hierarchical agglomerative clustering groups a page's roughly 750 patch vectors into 32 representative centroids at indexing time. Tencent stresses that this requires no training — it is a plug-in step — and quotes the result as 3.81 GiB per million pages, which is the figure that makes a document archive of this kind practical on ordinary hardware. Accuracy holds up. The model reaches 66.02 nDCG@10 on ViDoRe V3, 92.07 on V1 and 73.38 on V2, second only to its own 8B teacher in the vendor's table and ahead of every third-party model listed, including 8-billion-parameter competitors. It was distilled from that teacher using a recipe Tencent calls anchor-preserving relation distillation. Two cautions for the reader. The weight files hold 4.54 billion parameters, while the vendor's comparison table reports 4.61 billion — Tencent does not explain the gap. And the backbone is not Tencent's own: the table cites Qwen3.5-4B, Alibaba's model, while the Hugging Face metadata records the August preview EVIE-Preview-4.5B as the immediate base. Weights are Apache 2.0 with no regional carve-out, and no technical paper had appeared at release. Adoption in the 30 days to 16 September 2026: 969 downloads and 19 likes.

▸Videos

No videos yet.

▸Reviews