EVIE-8B vs EVIE-Preview-4.5B: A 1.39-Point Gain for 32× Wider Vectors

Three weeks after Tencent released EVIE-Preview-4.5B and called it the most accurate visual document retriever on the ViDoRe boards, Tencent released EVIE-8B and quietly moved the goalposts its own 128-dimensional model had been standing on. EVIE-8B is the 8.41B flagship teacher with 4096-dimensional per-token embeddings; EVIE-Preview-4.5B is the 4.54B compact retriever whose whole pitch was that 128 dimensions were enough to beat models four times its size. On the refreshed leaderboard inside the EVIE-8B card, EVIE-Preview-4.5B now sits in third place within its own family — behind the new EVIE-8B and behind EVIE-4.5B, a student Tencent shipped the same morning. If you are choosing which of the two named models to index a document corpus against, the numbers on Tencent's cards say the 8B is about 1.39 points better on ViDoRe V3 and 3.36 points better on ViDoRe V2 — and that it costs 32 times more index space per vector to get there. This piece is about whether that trade is worth making, and it starts from the honest caveat that both scorecards are Tencent's own, and neither has been independently reproduced.

The short version

• Pick EVIE-8B if retrieval accuracy is the bottleneck and your corpus is small enough that raw index size does not matter — you are measuring in thousands of pages, not millions, and you want the best open retriever Tencent has published.

• Pick EVIE-Preview-4.5B if index cost, per-page latency, or GPU budget is a real constraint — its 128D vectors are the entire reason it exists, and the accuracy gap to the 8B is small on ViDoRe V1 and modest on V3.

• Re-check both against EVIE-4.5B before you commit: Tencent released a same-day student with runtime-truncatable vectors and token compression that undercuts both on the efficiency frontier, and any comparison that ignores it is already dated.

• Treat every number here as vendor-reported. EVIE-8B has not been run by anyone outside Tencent; EVIE-Preview-4.5B's figures come from Tencent's own reproduction scripts.

Where they actually differ

• Params — EVIE-8B: 8.41B. EVIE-Preview-4.5B: 4.54B.

• Backbone — Qwen3.5-9B with full bidirectional attention vs Qwen3.5-4B with interleaved GatedDeltaNet linear attention and full attention.

• Embedding — 4096-dimensional per-token multi-vector vs native 128-dimensional per-token multi-vector, both scored with MaxSim late interaction.

• ViDoRe V1 (10 tasks, nDCG@5) — 92.18 vs 91.73.

• ViDoRe V2 (4 tasks, nDCG@5) — 74.23 vs 70.87. This is the biggest relative gap.

• ViDoRe V3 (48 tasks, nDCG@10) — 66.75 vs 65.36.

• Visual-token budget behind those headline numbers — 1,024 tokens/page for EVIE-8B's eval vs 1,792 tokens/page for EVIE-Preview-4.5B's headline tier (at its 768-token training tier the preview posts 64.56).

• Raw BF16 index per 1M pages — not published for EVIE-8B vs 179.2 GiB at 768 tokens/page and 420.5 GiB at 1,792 for the preview.

• License and release — Apache-2.0, quiet ship 2026-09-04 vs Apache-2.0, quiet ship 2026-08-17.

The scoreboard above restates the same portrait. Keep two caveats in view while you read it: the EVIE-8B column and the EVIE-Preview-4.5B column come from two different Tencent model cards, and the 8B's headline V3 number is measured at a lower per-page visual-token budget than the preview's headline number.

Two Tencent cards, talking to each other

The honest framing for this matchup is that it is a family argument, not an independent contest. EVIE-Preview-4.5B's own card, published August 17, claims "Rank #1 on ViDoRe V3" at 65.36 nDCG@10, beating webAI-ColVec1.1-8b by 0.04. EVIE-8B's card, published September 4, re-lists that same 65.36 as the third-best number on the page, below EVIE-8B's 66.75 and EVIE-4.5B's 66.02, and shows the 8B winning all eight public ViDoRe V3 domains against the preview. Both scorecards were produced with Tencent's own evaluation harness, and neither model has an independent run anywhere in the literature yet. So the comparison you are reading is Tencent's account of Tencent beating Tencent — internally consistent, plausibly engineered that way on purpose, and unverified by anyone with no stake in the outcome.

The tencent/EVIE-8B card above is the challenger's public surface: an 8.41B flagship teacher with 4096D embeddings and the family's refreshed ViDoRe table at the top.

The tencent/EVIE-Preview-4.5B card above is the incumbent's: a 4.54B retriever whose native 128D vectors and published index-cost math are still its defining feature.

The 1.39-point question

The raw gap on ViDoRe V3 is small — 1.39 nDCG@10 points between EVIE-8B's 66.75 and EVIE-Preview-4.5B's 65.36 — and it comes with a token-budget asterisk that matters for deployment. EVIE-8B's evaluation protocol caps documents at 1,024 visual tokens per page; the preview needed 1,792 tokens per page to post its headline 65.36, and scored only 64.56 at its own 768-token training budget. On those numbers the 8B is not merely more accurate at a comparable budget — it is more accurate at a smaller one, which is the direction you want for per-page latency. The bigger relative win is on ViDoRe V2, where EVIE-8B reports 74.23 against the preview's 70.87, a 3.36-point jump on the board that includes the harder synthesized document tasks. ViDoRe V1, the oldest and most saturated board, barely moves: 92.18 vs 91.73. That pattern — flat where everyone is already near the ceiling, decisive where the tasks are newer and harder — is the profile of a genuine capability increase rather than leaderboard noise, but it is still Tencent's own measurement of it.

The index is the real opponent

Accuracy is only half this decision, because the two models make radically different promises about what you store. EVIE-Preview-4.5B's reason to exist is its native 128-dimensional token vector: the card publishes exact index math (179.2 GiB of raw BF16 index per million pages at its training budget, 420.5 GiB at the extrapolated tier) and points out that a narrower vector cuts storage and MaxSim scoring work in direct proportion. EVIE-8B's token vectors are 4096-dimensional — 32 times wider per vector — and the EVIE-8B card publishes no index-size figure at all. That omission is itself information: at the same per-page token count, a raw BF16 EVIE-8B index is on the order of 32× the preview's, which puts a million-page corpus in the multi-terabyte range before quantization and makes the flagship something you point at a few hundred thousand pages, not something you casually host at scale. The flagship's width buys retrieval fidelity; it does not buy deployability.

The third option Tencent shipped the same day

No honest version of this comparison stops at the two named models, because the release that contained EVIE-8B also contained EVIE-4.5B — a 4.61B student distilled from the 8B teacher, with a single 2048-dimensional projection that truncates at runtime to 64, 128, 256, 512, 1024, or 2048 dimensions, plus a training-free token-compression pass Tencent calls HAC that clusters a page's visual tokens down to 32–64 vectors and, per the card, an index footprint of 3.81 GiB per million pages. On Tencent's own table EVIE-4.5B scores 66.02 on ViDoRe V3 — only 0.73 behind the 8B teacher and 0.66 ahead of EVIE-Preview-4.5B — which makes it the answer to the exact dilemma this matchup poses. If what you want is "most of the 8B's accuracy without the 8B's index," Tencent has already shipped that model, and it is neither of the two this page is titled after. EVIE-Preview-4.5B's remaining case is narrow but real: it is the checkpoint already in production for anyone who deployed it in August, and its fixed 128D profile is simpler to reason about than a matryoshka that asks you to choose your dimension at runtime.

Which should you index with

Match the model to the constraint that actually binds:

• Index at EVIE-8B if you are building the accuracy reference point — a benchmark, a high-value legal or scientific corpus in the hundreds of thousands of pages, or a system where retrieval misses are expensive and storage is cheap. You are paying for the top of the family's curve, and the family intends the 4.5B student to be the thing you scale out later.

• Stay on EVIE-Preview-4.5B if you already indexed with it and the measured quality is acceptable — a re-index is a real cost, and the V3 gain you would chase is 1.39 points on a vendor card that no one has independently confirmed.

• Evaluate EVIE-4.5B before either if you are starting fresh at meaningful scale. The Prefix-MRL truncation and HAC compression are aimed directly at the storage arithmetic above, and Tencent's own numbers put it within striking distance of the teacher.

None of the three has a hosted API on any platform, including OrcaRouter, so the retrieval stage is a self-hosted decision either way. The stage after it does not have to be. A router like the one OrcaRouter runs across 200+ models keeps the model that reads your retrieved pages and writes the answer behind a single API with automatic failover across providers and provider list prices passed through at 0% markup — which is precisely the setup you want when the retriever feeding it is a three-day-old checkpoint with unverified claims. Build the index with the retriever that fits your storage, and keep the rest of the pipeline swappable until someone outside Tencent confirms which of these scorecards is real.