The hub

Models you can run in this tab

Weights, a manifest that says what they are, and a sample the model has to reproduce. Nothing below is a screenshot. The list is fetched from the catalog when this page loads, and the button downloads a real model into your browser and checks it.

Verify one, here

This runs the whole path: ask the browser what it can do → fetch the manifest → download the weights → hash them → build the architecture → run the sample input → compare with the sample output. The order matters — checking the environment last would mean telling you it will not work after 10 MB had arrived, and hashing after loading would mean the wrong bytes are already in memory.

Download and verify

idle

It really downloads. The smallest model in the catalog is chosen so this costs about ten megabytes.

The catalog

Two kinds of thing live here, and the difference is worth reading: trained-by-borch means the weights were produced by this library, in a browser. converted-from-torch means somebody else trained them and they were carried across — the registry keeps a provenance note for each saying whose they were.

Two licences, not one. The weights carry theirs and the data they were trained on carries its own, and they are rarely the same sentence — Apache-2.0 weights trained on ImageNet do not make ImageNet's terms go away.

In your own page

import { init } from "borch-ts";
import { load, verify } from "borch-hub";

await init();
const { model, manifest } = await load(MANIFEST_URL);
const badge = await verify(model, manifest, MANIFEST_URL);

load takes a fetch, so a private model, an organisation account or a signed URL all go through that one hook. The catalog itself is a single document — index.json — and this page has no list of its own; it reads that one.

What this is not

There is no torchvision-shaped name for it. torch.hub is the closest and it fetches models from repositories without a catalog of its own. The three things here — a published manifest, a hash checked before loading, and an environment check before downloading — are in neither timm nor torch.hub. The architectures come from bimm, which is timm's seat; the hub knows the catalog and the catalog does not know the hub.