EVIE-8B: Tencent Quietly Ships Its Most Accurate Visual Document Retriever Yet

Three weeks ago, Tencent's EVIE-Preview-4.5B was the most accurate open visual document retriever on the ViDoRe boards, on the strength of figures Tencent reported itself. On September 4, Tencent quietly replaced that model at the top of its own leaderboard — not with an outside challenger, but with a bigger sibling released the same morning. EVIE-8B, an 8.41B-parameter late-interaction retriever on a Qwen3.5-9B backbone with 4096-dimensional per-token embeddings, landed as weights plus code on Hugging Face and GitHub with no announcement, no blog post, and no paper. A third checkpoint, EVIE-4.5B, appeared in the same minute and now sits between the two. Every claim attached to this release — the scores, the architecture, even what the models are for — currently rests on what Tencent's own repositories say, because no one outside Tencent has published anything about EVIE-8B yet. This is what is actually knowable from the repo, and what is not yet confirmed.

What Tencent shipped on September 4

On 2026-09-04 two new repositories appeared in the tencent organization on Hugging Face roughly a minute apart: tencent/EVIE-8B and tencent/EVIE-4.5B. A consolidated GitHub repository, Tencent/EVIE, went up alongside them carrying the training code, the ColQwen3.5 inference engine, and the 138-task evaluation harness. Both model repos are Apache-2.0, both use the colpali-engine library, and both carry the visual-document-retrieval pipeline tag. The older EVIE-Preview-4.5B from August 17 remains on Hugging Face under the "Preview" name, but it is no longer the family's flagship: in the refreshed leaderboard inside the EVIE-8B card, Tencent lists three of its own checkpoints on top — EVIE-8B at 66.75 ViDoRe V3 nDCG@10, EVIE-4.5B at 66.02, and EVIE-Preview-4.5B at 65.36.

There has been no announcement anywhere. The EVIE-8B card's release note says the weights, inference pipelines, and evaluation suites are open-sourced, then adds that "full technical details, architectural ablations, and the formal research paper will be updated in an upcoming release." The GitHub repository had four commits and no releases at the time of writing. Neither new repo shows meaningful adoption yet — no community forks with independent results, no third-party evaluations, and download counters still near zero two days after the repos went public. On the evidence, this is a ship-the-weights-first, write-the-paper-later release from Tencent, in the same quiet pattern the lab used for EVIE-Preview-4.5B and UI-Mate-9B in August.

The tencent/EVIE-8B model card above is the release's public surface: a visual-document-retrieval checkpoint tagged for the colpali-engine library, Apache-2.0 licensed, with a 4096-dimensional late-interaction embedding and the family's own refreshed ViDoRe table near the top of the card.

EVIE-8B, the flagship teacher

Architecturally, EVIE-8B belongs to the ColPali/ColBERT late-interaction family. Instead of crushing a page into one embedding vector — the lossy move that single-vector dense retrieval makes — it runs a vision-language backbone over the page image and keeps a per-token multi-vector representation, then scores relevance with the MaxSim operator: each query token takes its best match across the page's tokens, and those maxima sum into the relevance score. EVIE-8B's specific choices are a Qwen3.5-9B backbone run with full bidirectional attention (the model is encoder-ized, so the causal mask is switched off across the vision-text sequence) and a wide 4096-dimensional projection per token. The point of the width is fidelity: spatial layout, typography, chart structure, and tabular relations survive in the representation instead of being averaged away.

Tencent positions EVIE-8B explicitly as the "flagship teacher." The EVIE-4.5B student is trained from it with a recipe Tencent calls EVIE-ARD — anchor-preserving, capacity-aware relation distillation that transfers token-relation topology and uses hard-negative margin supervision. In other words, the 8B model has two jobs: to set the accuracy ceiling for the product line, and to be the quality anchor that the compact student distills down from. On the training side, the card reports 775,635 document-query pairs, with mined hard negatives re-judged by a multimodal judge that promotes answerable candidates to positives, masks ambiguous ones out of the loss, and keeps only strictly irrelevant pages as true negatives. The released "a40" checkpoint is itself a weight-space mix, at α = 0.40, of two independently trained arms.

The scoreboard Tencent wrote for itself

All of the numbers below are vendor-reported. They come from Tencent's own model card and were produced with the evaluation harness Tencent ships in its repository; none has been independently reproduced, and EVIE-8B has not yet been run by anyone outside the lab.

• Params / backbone — 8.41B, Qwen3.5-9B, bidirectional full attention.

• Embedding — 4096-dimensional per-token multi-vector, MaxSim late interaction.

• ViDoRe V1 (10 tasks, nDCG@5) — 92.18.

• ViDoRe V2 (4 tasks, nDCG@5) — 74.23.

• ViDoRe V3 (48 tasks, nDCG@10) — 66.75, with a per-domain sweep the card claims wins all eight public domains.

• JinaVDR (76 tasks, nDCG@10) — 83.30; four-board macro average nDCG@10 — 79.51 across the 138-task suite.

• License — Apache-2.0; weights, inference, and eval code open.

The most interesting row in Tencent's table is the one where Tencent beats Tencent. The EVIE-Preview-4.5B card from August claimed "Rank #1 on ViDoRe V3" at 65.36 nDCG@10, edging webAI-ColVec1.1-8b's 65.32. In the refreshed table inside the EVIE-8B card, that same 65.36 is now third place overall, behind EVIE-8B's 66.75 and EVIE-4.5B's 66.02. EVIE-8B's per-domain breakdown beats the old preview in all eight public ViDoRe V3 domains — CompSci 81.86 vs 80.65, Finance EN 71.23 vs 70.50, Pharma 70.81 vs 69.20, and so on — for an average gain of 1.39 points. Whether that gain survives an independent run is exactly the open question; but the internal consistency is worth noting, because a teacher that could not beat its own student would be a strange thing to ship.

The family graphic above restates the three-checkpoint ranking Tencent published in the EVIE-8B card on September 4 — EVIE-8B at 66.75, EVIE-4.5B at 66.02, EVIE-Preview-4.5B at 65.36 on ViDoRe V3 nDCG@10 — with the 128D preview giving way to the 4096D teacher and the Prefix-MRL student.

Why the flagship is not the deployment model

There is a reason the accuracy leader is called a teacher rather than a default. Width costs storage and scoring compute, and 4096 dimensions per token is very wide. In BF16, one EVIE-8B token vector is 8 KiB before any quantization, and a dense page can produce hundreds of token-vectors — the card's own evaluation protocol caps documents at 1,024 visual tokens per page. Tencent publishes no index-size figure for EVIE-8B, which is conspicuous because the EVIE-Preview-4.5B card published exact ones for its 128-dimensional index: 179.2 GiB of raw BF16 index per million pages at its 768-token training budget, 420.5 GiB at the extrapolated 1,792-token tier. Vector-for-vector, EVIE-8B's embeddings are 32 times wider than the preview's, so a raw EVIE-8B index at the same token budget runs to the multi-terabyte range per million pages — before the retrieval system that must hold it in RAM has even entered the conversation.

The same-day release of EVIE-4.5B is the tell for how Tencent expects that tension to be resolved. EVIE-4.5B is the 4.61B student with a single 2048-dimensional projection that truncates at runtime to 64, 128, 256, 512, 1024, or 2048 dimensions (Tencent calls the scheme Prefix-MRL), plus a training-free token-compression pass it calls HAC that clusters a page's ~750 visual tokens down to 32–64 vectors. Tencent's headline for that model is an index footprint of 3.81 GiB per million pages — a number that, like the rest, is vendor-reported. Read together, the release looks less like "one new 8B model" and more like a deliberate family: the 8B sets the ceiling and trains the student, and the student is the one sized to actually sit behind a production index.

How to run EVIE-8B today

EVIE-8B ships as weights only. There is no hosted API on any platform — including OrcaRouter — so the only route today is to self-host the checkpoint and run it through the ColQwen3.5 path in colpali-engine. The card's quick start is short: load the model in BF16 with flash-attention, call enable_bidirectional_attention(), embed the page images and the text query, then score with the processor's MaxSim. The bidirectional step is not optional decoration — released colpali-engine builds ColQwen3.5 with causal masks by default, and the preview model's card documented that leaving the mask in place costs roughly 1.1 points on top-hit MaxSim. The same helper ships for EVIE-8B. An 8.41B BF16 checkpoint is on the order of 17 GB of weights by the two-bytes-per-parameter rule of thumb, so this is a single-GPU-with-headroom proposition rather than an edge model, and the corpus lives wherever you keep the index, not on the model.

What EVIE-8B changes for a document RAG stack

For anyone building retrieval over scans, filings, charts, or forms, the practical change is that the quality ceiling for open visual document retrieval just moved, and it moved inside a family that now spans three price-performance points. A retriever only gets you the pages, though. The model that reads those pages and writes the answer is a separate decision, and that is the layer where routing earns its keep: OrcaRouter exposes one API across 200+ models at provider list price with 0% markup, so the reader model behind your retriever can be swapped without a rewrite, and a vendor price cut on any routed model is live the same day it is announced. That separation matters more when the retrieval half is a brand-new, unproven checkpoint — you can adopt EVIE-8B for indexing while keeping the generation layer behind an interface that does not couple you to any single model.

What to watch next

Three things would turn this from a quiet repo drop into a settled story. First, an independent run: until someone outside Tencent reproduces the ViDoRe and JinaVDR numbers, both EVIE-8B's "rank #1" and the old preview's "rank #1" are self-reported claims that happen to be in tension with each other. Second, the promised paper and ablations — the 4096-dimensional choice, the weight-mix checkpoint, and the distillation recipe are all currently assertions without the write-up to back them. Third, the naming: with EVIE-4.5B now occupying the spot EVIE-Preview-4.5B held three weeks ago, it is an open question whether Tencent keeps the "Preview" line alive or folds it into the renamed family. Given the cadence — three visual-document-retrieval checkpoints in three weeks — the more useful assumption is that this line is being invested in hard, and that the quiet releases are the lab's way of shipping ahead of the paper.