IT EN
Start a trial

Operational notes Observatory

Tencent’s EVIE-Preview-4.5B: the model weighs 8.5 GB. Indexing your documents doesn’t

6 min read

Row of cylindrical metal industrial silos photographed close-up in black and white, with a dark building in the background
The model fits in a single 8.5-gigabyte container. What you put into it fills the silos next to it, one after another.

On 17 August 2026, at 11:53 UTC, Tencent published EVIE-Preview-4.5B on Hugging Face: not a model that generates text, but an engine that reads a document page as an image and returns which pages, out of millions, answer a query — without going through OCR. For an organisation that needs to search archives that were never born digital — contracts, personnel files, scanned financial tables, hand-filled forms — this is exactly the kind of component that ends up in a tender. Before it does, two sums need doing: what it actually costs to run, and what the card declares, or leaves unsaid, about the data it was trained on. We checked, as of 24 August.

The facts, in order

  • Release: repository created on 17 August 2026 at 11:53:54 UTC, with a minor update on 18 August. Organisation tencent, verification badge active (isVerified: true on the public organisations endpoint).
  • Access: gated: false. Checked without authenticating: the request for the configuration file returns a direct redirect to the file, no login required.
  • Adoption, today: 337 downloads, 39 likes — the numbers of a model published a week ago, not of a release already filtered by the market.
  • Weight: a single weights file, 4,539,593,344 parameters in BF16, about 8.5 GB — no shards, no official quantisation (GGUF, FP8) published yet.
  • Base: built on Alibaba’s Qwen/Qwen3.5-4B, itself under Apache 2.0 and not gated.
  • Licence: Apache 2.0, as an open-weights licence should read — a LICENSE.txt file is present and consistent with the declared tag, with the caveat, written into the text itself, that third-party components remain under their own original licences.
  • Documentation: no technical report, no arXiv identifier. Everything we know about the model sits in a 396-line README.
  • The link the card promises, github.com/Tencent/EVIE-Preview-4.5B, exists and resolves: not a promise left without a page.

What it actually is

EVIE does not answer questions: it returns a relevance score for every page. Each page — seen as an image, not transcribed — becomes a set of 128-dimensional vectors, one per visual token; the query becomes another set of vectors; the score compares the two sets token by token, a technique known as late interaction, the same family as ColPali. Useful exactly where OCR fails: tables, forms, stamps, poor scans. The answer to the question, though, is generated by a separate downstream model that reads the retrieved pages: EVIE finds, it does not explain.

The hardware that’s missing, and the number that isn’t

The card states no minimum configuration: no recommended GPU, no memory threshold. The only figure that allows an estimate is the file’s weight, 8.5 GB — plausibly compatible with a single professional card. That is our own inference, not a claim made by the card, and it should be treated as such.

The figure the card does declare concerns a different line item, the one that actually decides whether the project stands up: the index. The documentation spells out the formula — 1,763.58 vectors per page, 128 dimensions, two bytes per value — and the result: 179.2 GiB per million pages at the training tier (768 visual tokens per page), 420.5 GiB per million pages at the extended tier. On the latter the card is candid to the end: “The 1,792 tier is pure test-time extrapolation — the same weights, never trained or fine-tuned at that budget, and never re-exported.” Across ten million pages — an average public-sector historical archive, not an edge case — that is 1.8 terabytes at the light tier and 4.1 at the heavy one. Nothing on the card says what it takes to query that index within a useful time: no vector database, no latency figure at that scale. Downloadable weights without the compute that makes them runnable are not sovereignty: here the missing arithmetic isn’t the GPU’s — already measured elsewhere for a model ten times heavier — it is the space in which the client’s archive, turned into vectors, has to sit still, be fast to read, and stay inside the same perimeter as the original documents.

The training data: two declared datasets, one missing

The card’s metadata declares two datasets: vidore/vidore_benchmark and vidore/vidore_benchmark_v2. Read in full, the README makes clear that those two names are not the training material: they are the evaluation boards, the same twenty-two public sets that the reproduce.sh script downloads to reproduce the numbers in the table. The material the model actually learned from is described elsewhere, in a paragraph a few lines long: “EVIE was trained on approximately 0.8 million high-quality image-query pairs spanning multilingual documents, technical reports, complex financial tables, infographics and document visual QA.” No source named, no licence, no date range. The same point we made about a technical report far longer than this one applies here: the paragraph on data is shorter than the one on installing via pip. For anyone evaluating the model — a Chinese organisation, with everything that implies for provenance and governance — the question is not rhetorical: if those pairs were drawn from real documents gathered elsewhere, a piece of somebody else’s perimeter is already inside the weights you are about to install in your own.

What we don’t know

We have not tested the model on our own documents. The declared top ranking — first on ViDoRe V3 and on V1+V2 — is computed with the vendor’s own script for its own row in the table; the competitors’ rows are, by the README’s own admission, scores published by each vendor, not an independent head-to-head test. We offer no legal qualification of the licence or of the data’s provenance. One minor detail, checked today: the README states that Sentence Transformers 6.0.0 is “not on PyPI yet”; as of 24 August, it already is. A week-old README ages faster than it looks — one more reason to check rather than trust.

How we solve this

Verifying a model like this does not stop at the licence: it enters the model registry as a dated line — version, file digest, channel and date of acquisition, the licence text archived as it read on the day you checked it, approved uses and who approved them — ready to put in front of an inspection, a tender, or a board. The same holds for the index that comes out of it: it is a derivative of your documents, and in our set-up it stays inside the same perimeter, never on third-party storage chosen for convenience. The same system ties together models, data, contracts, archives and documents into a single operating model, on which AI agents execute decisions with a human operator in command — for large enterprises, defence, government and healthcare. On-premise, on self-contained machines that need no deep integration into the client’s network, or on a dedicated cloud with a dedicated VPN and a data centre in Italy — always with shared management.

See the service · Talk to an engineer

Do you have an archive of scanned documents that nobody can search well? Half an hour with one of our engineers, free of charge: we bring the real estimate of space and infrastructure before a model like this goes into a tender.

Sources