EVIE-8B 2026: Tencent ViDoRe document retrieval and vendor evals
Tencent opened two EVIE retriever checkpoints. Separate the 8.41B / 4.61B SKUs, the vendor ViDoRe table, and the compressed-row drop.
Around 7 September 2026 Tencent’s IMA Product Center and Youtu Lab released EVIE-8B and EVIE-4.5B: retrievers that search scanned pages, tables, and slides as images instead of OCR’d text. The story is not another chat flagship. It is the lab’s own ViDoRe table — and the fact that the compressed SKU and the headline score are different rows.
The query is EVIE-8B and ViDoRe V3. This page follows the Hugging Face cards tencent/EVIE-8B / tencent/EVIE-4.5B, GitHub Tencent/EVIE, and DataNorth’s 7 September recap. The scores sit on a vendor protocol; do not relabel them as an independent lab rerun. Tencent’s other open drop that month: Hy4 preview.
What the lab said it shipped
EVIE is Evidence-Vector-Informed Embedding. Both SKUs use ColQwen3.5 late interaction: hundreds of vectors per page, scored with MaxSim, not one summary vector. The 8B is the teacher (bidirectional attention, 4096-D); the 4.5B is the student (single-projection Prefix-MRL). The card names Zifei Wang (IMA) and Wei Wen (Youtu, corresponding). Training lists 775,635 document–query pairs; the a40 checkpoint is a weight-space mix of two arms (α=0.40). A formal paper is marked upcoming. Same base family: Qwen3.8-27B.
How to read the three numbers
- 1
Name the SKU before 66.75
The 8B sample uses colpali-engine or Sentence Transformers’ MultiVectorEncoder with FlashAttention and bidirectional attention. The 4.5B can truncate 2048-D to 1024, 512, 256, 128, or 64 at query time without a second checkpoint. The card prints 65.27 on V3 at 128-D, down from 66.02. Licence: Apache 2.0 for weights and companion code.
- 2
66.75 is a custom vendor protocol
V3 is eight domains × six query languages. The 8B average is 66.75, from 81.86 in computer science to 52.11 in physics. Named rivals: webAI-ColVec1.1-8b 65.32, VultronRetrieverPrime-8B 64.26, nemotron-colembed-vl-8b-v2 63.54. On V1 nDCG@5 NVIDIA’s row is 92.65 versus EVIE-8B’s 92.18. The crown sits on the newer V2/V3 slices.
- 3
3.81 GiB maps to a different score
HAC needs no extra training and folds ~750 page vectors to 32 or 64. The card’s default compact setting is d64 K32 at 3.81 GiB per million pages. Same table: uncompressed 4.5B V3 is 66.02; compressed is 59.58, slightly behind tomoro-colqwen3-embed-4b at 60.16. Both rows are Tencent’s. Do not quote only the highlight.
Late interaction trades storage, not magic
Hundreds of vectors keep tables and layout; they also grow the index. Prefix-MRL cuts width; HAC cuts count. The card itself treats the compact and full-width scores as different settings.
138 tasks are not an AA index
The four metric families are nDCG, Recall, MAP, and MRR. Macro averages sit on the card — e.g. 79.51 nDCG@10 across four boards for the 8B. Do not resticker that as an Artificial Analysis Intelligence Index.
Paper and memory bills are still pending
The announcement says weights, inference, and the eval suite are open; architectural ablations and the formal paper are “upcoming.” Speed and peak VRAM are not a deployment invoice on the card.
| Claim | Checkable source | How to read it |
|---|---|---|
| 8.41B / 4.61B, Apache 2.0 | HF cards · GitHub Tencent/EVIE | Architecture from the official table |
| V3 66.75 / 66.02 | Card ViDoRe comparison table | Vendor protocol, not an independent rerun |
| 3.81 GiB and 59.58 | 4.5B card HAC / Prefix-MRL rows | Compressed row is separate from the headline |
Tencent’s card: full technical details, architectural ablations, and the formal paper will be updated in an upcoming release. It does not promise that rival scores were rerun on the same protocol box.
# EVIE · Tencent IMA / Youtu · as of 2026-09
sku_8b: 8.41B on Qwen3.5-9B, 4096D
sku_45: 4.61B on Qwen3.5-4B, Prefix-MRL 64–2048D
license: Apache-2.0
hf: tencent/EVIE-8B · tencent/EVIE-4.5B
vidore_v3_ndcg10: 66.75 / 66.02
protocol: paired-all-pages-dedup+process_queries+ndcg2r-20260827
hac_index: 3.81 GiB / 1M pages (d64 K32)
compressed_v3: 59.58 (4.5B d64 K32)Boundaries
- Vendor ViDoRe ≠ third-party index
- 66.75 / 66.02 hang on a named custom protocol. Do not relabel them as an AA Intelligence Index, and do not crown them against the same week’s chat-flagship vendor tables.
- Apache 2.0 ≠ 66.02 after compression
- The licence covers commercial self-hosting. The 3.81 GiB row is 59.58. Dense financial tables lose more to clustering than text-heavy reports.
- This page is not a meeting manual
- It only explains the release and the evals. For a short huddle with a board, create a room on wbmeet; see how to create and join a room.
Questions worth checking
Is EVIE another chat LLM?
No. It is a visual document retriever: text query, page images, ranked pages out. Do not stack it on the same month’s Hy4 preview weights.
Does 66.75 mean it beats every open retriever?
Not by itself. It is a V3 average under the lab’s protocol. NVIDIA’s nemotron-colembed-vl-8b-v2 is higher on V1. Same-box reruns of rivals are not claimed.
Is 3.81 GiB the official default deploy?
The card labels d64 K32 as the default compact setting and prints 66.02 on the uncompressed row. Measure both on your own pages.
Is the technical report out?
As of 2026-09-09 the card still says the paper and ablations are upcoming. Checkable pieces are the repos, the protocol name, and the in-table figures.