Tencent pushed weights for two visual document retrieval models to Hugging Face on 7 September under Apache-2.0, with no announcement and no coverage at the time of writing. The release commit — "EVIE-8B: teacher weights, inference, 138-task eval" — is timestamped 09:04:08 UTC, or 17:04 in Beijing, with documentation commits continuing to 10:41 UTC.
What shipped
EVIE-8B is an 8.41B-parameter model on a Qwen3.5-9B base producing 4096-dimensional per-token embeddings; EVIE-4.5B is a 4.61B distilled student with truncatable dimensions from 64 to 2048. Both are late-interaction retrievers in the ColBERT lineage — they keep per-token vectors and score with MaxSim rather than compressing a page into one vector, which is what preserves table and layout structure. The 8B repo is 16.84 GB across 47 files; the 4.5B is 9.11 GB across 88, and includes training code.
The claim its own table does not support
The card leads with two gold Rank #1 badges. The first is defensible: 66.75 nDCG@10 on ViDoRe V3 does top the comparison table, ahead of webAI-ColVec1.1-8b at 65.32 and nemotron-colembed-vl-8b-v2 at 63.54. The second badge reads "ViDoRe V1+V2 — 92.18 · Rank #1", and 92.18 is EVIE-8B's ViDoRe V1 score. In the same table, nemotron-colembed-vl-8b-v2 scores 92.65 and VultronRetrieverCore-4.5B scores 92.21 on that board. On V1, Tencent's own numbers place EVIE-8B third.
What the benchmark framing hides
The card states that the released checkpoint is an "a40" weight-space mix at alpha = 0.40 of two independently trained arms, selected because it achieves the peak V3 score. A blend ratio chosen for benchmark performance is a benchmark-fitted artefact, and the card does not report how the unblended arms score. There is also no paper: the release note says the formal write-up and architectural ablations "will be updated in an upcoming release". Every number here is self-reported on Tencent's own card, not an independent leaderboard submission.
Not to be confused with August
Tencent published EVIE-Preview-4.5B on 17 August, a distinct 128-dimension checkpoint that appears in this card's table at 65.36. The repositories for the two models released today were created on 4 September as empty placeholders; the weights themselves did not exist publicly until Monday.