Every model SecureRAG downloads, listed with its size and licence

Five model files exist in total, all of them open-licensed and small. Two cover almost everyone at roughly 25 MB each, and the two generative models are optional downloads you never have to press. This page gives the numbers.

Last updated:

Nothing here is fetched when a page loads. A model file is downloaded once, after you press a button that names the file and its size, and it is stored in your own browser cache. After that the whole application works offline.

All five model files in one table

Table 1 — every model SecureRAG can download
ModelParametersQuantisationFirst downloadVector dimensionsLanguagesLicenceSource
bge-small-zh-v1.5 — default for Chinese≈ 24 M (listed as 0.02 B on the model page)int8 dynamic quantisation, ONNX≈ 25 MB512Chinese; short English text works but retrieves worseMIT, as stated on the model pagehttps://huggingface.co/Xenova/bge-small-zh-v1.5
all-MiniLM-L6-v2 — default for English≈ 23 Mint8 dynamic quantisation, ONNX≈ 23 MB384EnglishApache-2.0, as stated on the model pagehttps://huggingface.co/Xenova/all-MiniLM-L6-v2
multilingual-e5-small — optional, mixed-language documents≈ 118 M (0.1 B on the model page)int8 dynamic quantisation, ONNX≈ 120 MB384Over 100 languages; the pick when one library mixes Chinese and EnglishMIT, as stated on the model pagehttps://huggingface.co/Xenova/multilingual-e5-small
Qwen2.5-0.5B-Instruct — optional, writes answers0.49 B total, 0.36 B excluding embeddings4-bit weight-only quantisation≈ 400 MBNot applicable; it produces text, not search vectors29+ languages, Chinese includedApache-2.0, as stated on the model pagehttps://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct
Qwen2.5-1.5B-Instruct — optional, writes answers1.54 B total, 1.31 B excluding embeddings4-bit weight-only quantisation≈ 1.0 GBNot applicable29+ languages, Chinese includedApache-2.0, as stated on the model pagehttps://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct
Why sizes differ from the original repositories: the figures above are for the quantised ONNX builds we actually serve, which are far smaller than the original training checkpoints. A full-precision bge-small-zh-v1.5, for instance, is several times heavier. Upstream parameter counts and licences can change; the model page always wins over this table, and we correct this page when a number turns out to be wrong.

Hardware, and what WebGPU changes

Embedding runs on a Web Worker, so the interface stays responsive while a document is indexed. When your browser exposes WebGPU, that worker runs the model on the graphics chip; when it does not, the same model runs through the WebAssembly build of onnxruntime that ships with the site, on the CPU. The CPU path is several times slower, and on a long PDF the difference is a few seconds against a minute. No feature is missing on the CPU path — it is the same embeddings and the same results.

Table 2 — what each setup can realistically do
SetupRetrieval and embeddingThe optional 1.5B generation tierWhat we suggest
Desktop with WebGPU (Chrome, Edge, recent Safari)Runs on the GPU, several times faster than the CPU pathTurns over a few words per second — readable, well behind a cloud chatbotThe intended setup. 300 MB of free memory is enough for retrieval only
Desktop without WebGPU (Firefox, older Safari, older GPUs)Falls back to the WebAssembly build bundled with the siteSlow. Fine for one short paragraph, tedious for a long answerUse the retrieval-only tier here; it is fully functional
Phone or tabletWorks, with lower collection limits and a memory ceiling the mobile browser enforcesOff by default: the download is larger than a mobile tab is usually allowed to holdGood for looking things up. Indexing a long report is better done on a computer
Browser without WebAssembly, or with storage for the site disabledNot supportedNot supportedWe say so on the first screen rather than fail halfway through indexing
Memory guidance is guidance, not a measurement. Plan on roughly 300 MB of free memory for retrieval-only work and about 1.5 GB for the largest generation tier, on top of whatever your other tabs are using. These are our recommendations for a smooth run, not tested minimums, because every browser and operating system accounts for memory differently.

The 30 MB rule, and how a download is confirmed

The default embedding models are under 30 MB and cover most libraries. Four things are heavier than that: the multilingual embedding model at about 120 MB, and the two generation models at roughly 400 MB and 1.0 GB. Any of the four is offered, never imposed, and each purchase of your bandwidth is confirmed twice: once when you open the download panel, once when you press the button.

01

The panel states the cost first

Before a single byte moves, the panel shows the model name, the exact file size, the host it will be fetched from and the licence it carries. Nothing is requested while you read it.

02

You press a button that names the size

The action reads like “Download 120 MB”, not “improve my results”. Closing the panel instead cancels the whole thing, and no request was sent.

03

Progress is visible, and cancelling is real

You see transferred bytes while the file arrives. Cancelling mid-download throws away the partial file rather than leaving a half model on disk.

04

The file is stored in your browser, and asked for once

It lands in Cache Storage under your browser profile. The same file is not requested again unless you clear site data, or the browser evicts it under storage pressure — in which case the panel appears again with the same figures.

What the request contains

  • A file name and a revision reference in the URL, and nothing else.
  • No query string carries your document, your question, your language or any identifier we could use to recognise you.
  • No cookie is attached, because there is no account system to set one and the request is made without credentials.
  • The model host sees your IP address, as any site you visit does. That is a property of the open internet, not something this application adds.
  • If the primary host is unreachable, the same file may be fetched from a mirror. The file name, the model and the size stay identical, and it is the same request shape.

Check the download figures yourself

  1. Open the tool, then open DevTools and go to the Network tab. Tick “Preserve log” and set the filter to all requests.
  2. Clear the request list, then add a document. Nothing new appears.
  3. Switch the model tier. Open the download panel and read the size before confirming — the number there is the number in Table 1.
  4. Press the download button and watch the single request: the path carries the model file name, the response is the file itself, and the transfer size rounds to the figure in the table.
  5. Go to the Application tab → Cache Storage. The model is there. Your document is not in that cache; chunks and vectors sit in IndexedDB.
One more check: set the Network tab to Offline after the download finishes and ask a question. It still answers, which shows the answering path never needed the network after the model arrived.
This page carries no advertising. Model and legal pages are kept free of ad slots on purpose. Ad behaviour on the rest of the site is described in the privacy policy, and no ad script loads before you agree to it.

If something here turns out to be wrong

Tell us at guweiicy@gmail.com, with the model name and the figure you checked. A licence or parameter count that contradicts the upstream model page is a mistake on our side, and we correct it and note the change on the changelog.