Published 7 September 2026
Tencent released EVIE-8B and EVIE-4.5B on 7 September 2026, two open-weight models that search document pages as pictures rather than as extracted text. Tencent reports 66.75 nDCG@10 for EVIE-8B on ViDoRe V3, the hardest of the three standard tests for this job, against 63.54 for NVIDIA’s nemotron-colembed-vl-8b-v2. Both models carry an Apache 2.0 licence, so you can use them commercially and run them on your own machines.
What do EVIE-8B and EVIE-4.5B do?
They find the right page. You ask a question, the model ranks a pile of document images, and it hands back the pages most likely to hold the answer. It never converts the page to text first. That matters for PDFs full of tables, charts, stamps and small print, where an OCR step usually destroys the layout that made the page readable.
EVIE stands for Evidence-Vector-Informed Embedding. Both models use late interaction, which means a page is stored as hundreds of small vectors instead of one summary vector. One vector per page throws away where things sat on the page. Hundreds of vectors keep it, and that is what lifts the scores on documents with real structure.
EVIE-8B is the larger model and was used to train the smaller one. Both are built on Qwen3.5, Alibaba’s open model family, wired up in the ColQwen3.5 configuration. A Tencent release standing on an Alibaba base is worth noticing. The Chinese open-weight labs now build on each other as readily as Western teams build on Llama.
EVIE-8B and EVIE-4.5B benchmark scores
The table below shows where the EVIE models land against the rivals Tencent chose for its own comparison. Read the last two rows before the first three.
| What is measured | EVIE | Named rival |
|---|---|---|
| ViDoRe V3, nDCG@10 (hardest page search) | EVIE-8B: 66.75 | webAI-ColVec1.1-8b: 65.32 |
| ViDoRe V2, nDCG@5 (unfamiliar domains) | EVIE-8B: 74.23 | VultronRetrieverPrime-8B: 68.18 |
| ViDoRe V1, nDCG@5 (the original test) | EVIE-8B: 92.18 | nemotron-colembed-vl-8b-v2: 92.65 |
| ViDoRe V3, nDCG@10, smaller model | EVIE-4.5B: 66.02 | VultronRetrieverCore-4.5B: 63.57 |
| ViDoRe V3 with storage compression on | EVIE-4.5B d64 K32: 59.58 | tomoro-colqwen3-embed-4b: 60.16 |
| Index storage per million pages | 3.81 GiB at 32 vectors a page | About 750 vectors a page uncompressed |
Every figure here is Tencent’s own. The company published a 138-task evaluation matrix for EVIE, but it does not say whether the rival scores in its leaderboard were rerun on the same setup or copied from published figures. The protocol named on the model card is a custom setting, so none of these numbers will line up with figures you find elsewhere unless you rerun them yourself. Nobody outside Tencent has checked the result.
How EVIE-4.5B cuts the storage bill
Late interaction has always had one weakness. A page becomes roughly 750 vectors, and a million pages becomes an index you have to justify to whoever signs off on storage. EVIE-4.5B attacks that from two directions, and this, not the benchmark table, is the interesting part of the release.
The first is Prefix-MRL. A single projection produces 2048 numbers per vector, and you can cut the tail off at query time down to 1024, 512, 256, 128 or 64 without loading a different model. At 128 numbers the ViDoRe V3 score falls from 66.02 to 65.27. A sixteenfold cut in width costs three quarters of a point.
The second is HAC, a clustering step that squeezes the 750 or so vectors of a page down to 32 or 64. It needs no training and can be switched on after you have already picked a model. Tencent’s smallest published setting stores a million pages in 3.81 GiB.
The practical details, all published by Tencent:
- Licence: Apache 2.0, covering both models and the training code
- Sizes: 8.41 billion parameters for EVIE-8B, 4.61 billion for EVIE-4.5B
- Base models: Qwen3.5-9B and Qwen3.5-4B, in the ColQwen3.5 configuration
- Embedding width: fixed at 4096 on EVIE-8B, 64 to 2048 on EVIE-4.5B
- Where to get it: Hugging Face, as tencent/EVIE-8B and tencent/EVIE-4.5B
- How to run it: colpali-engine, cloned from the Tencent/EVIE repository
What Tencent is not saying about EVIE
The compression and the headline score do not hold at the same time. The 66.02 for EVIE-4.5B is measured with every vector kept. The setting that actually delivers the advertised 3.81 GiB index scores 59.58, which is 6.4 points lower and slightly behind tomoro-colqwen3-embed-4b, a model Tencent lists as beaten. Both numbers sit in Tencent’s own tables, on different rows, and only one of them reaches the highlights.
Two smaller gaps. On ViDoRe V1, NVIDIA’s nemotron-colembed-vl-8b-v2 scores higher than EVIE-8B, so the claim of being most accurate rests on the two newer tests rather than all three. And there is no paper yet, no speed figures, and no statement of how much GPU memory either model needs. Tencent says the technical report is coming.
What this means
If you already run document search over PDFs, scanned reports, invoices or slide decks on ColPali or ColQwen, test EVIE-4.5B this week. It is a checkpoint swap rather than a rebuild, the licence lets you keep it, and it runs on your own hardware, which is the deciding factor for contracts and patient records that cannot leave the building. Do not test it for the 0.66 points it gains over Tencent’s own preview model from August. Test it because Prefix-MRL lets you change your mind about index size after you have indexed everything, and that flexibility is worth more than a fraction of a point.
Test the compressed setting first, not the full one. The gap between 66.02 and 59.58 is the whole decision, and it will land differently on your documents than on ViDoRe. Dense financial tables will suffer more from the clustering than text-heavy reports will. If you retrieve over clean text, or over an archive small enough that index size has never come up in a meeting, this release is safe to ignore.
For more information, visit the official announcement of EVIE-4.5B in the Tencent EVIE model card.
