How the Model Hub Is Built

What Is in the Published Library Right Now

Snapshot published 2026-09-18T19:23:58 under source policy model-hub-sources/1.1.0. These are counts of records that completed an import and passed validation, not an estimate of how much exists upstream.

The Sources It Reads

The hub reads a named list of sources. A source outside this list cannot reach a public page, and a field outside each source's allowlist is dropped before it reaches the catalog.

Weights and dataset files stay with their publishers. SAVRN stores metadata and links, copies no artifact, mirrors nothing, and never passes a gated download through a shared credential. Where a publisher gates a repository, the request for access is made to them, by you.

What Each Evidence Label Means

Evidence is a property of a claim, not of a page, so one page can carry several labels at once.

• Publisher reported: the publisher stated it on their own repository. Recording it is not agreement with it.
• Reported by a third party: somebody other than the publisher filed it against the repository. The reporter's name and link stay attached to the number.
• SAVRN reviewed: a named person at SAVRN checked the claim against its source and recorded the result.
• SAVRN measured: SAVRN ran the test itself, under a stated configuration. No model page currently carries this label.
• Unknown: the hub has no support for the claim. It is shown as unknown rather than filled in.

What a Missing Value Means

A blank field is a fact about the source, not a gap to be estimated. If a publisher did not state a revision, the page says the revision is unknown and no version page claims to be reproducible. If nobody reported a benchmark, no score is shown. If a parameter count is absent, the memory table is absent with it.

How Relationships Are Decided

A relationship is written only where a publisher declared it on their own repository. A paper that mentions a dataset is not a training disclosure, and the hub does not treat it as one. Where a declared target is not in the library, the relationship still shows the reference the publisher wrote, linked to its source.

A curator can suppress a declared relationship or correct its target or type, with a reason and a cited source. The publisher's declaration is kept as it was, the correction is marked on the page as a SAVRN correction, and it keeps the publisher's evidence label: correcting where a link points does not turn a publisher's disclosure into a SAVRN finding. When two records are the same thing, a curator can map one to the other; the duplicate stops being listed and its address leads to the record it duplicates. Variants, quantizations, adapters and mirrors are different records and stay listed separately.

• Derived from: derived_from
• Base model for: base_of
• Quantized from: quantized_from
• Adapter of: adapter_of
• Merged from: merged_from
• Trained on (disclosed): trained_on
• Evaluated on: evaluated_on
• Described by: described_by
• Published by: published_by
• Alternative to: alternative_to

How Access Is Described

• Available from the publisher
• Access requested at the publisher
• No download offered
• Withdrawn upstream
• Unknown

These are separate from the license. A repository can be open to download and still carry terms restricting what you may do with it, and the hub never collapses the two into a single "open" badge.

How Often It Refreshes, and What Happens When a Refresh Fails

Metadata refreshes on a schedule. Each run walks a named scope, checkpoints after every page, and records what it saw. A run that fails part way through keeps everything it already wrote and stores its position, so the next run continues rather than starting over.

A failed or partial run never empties the library. The published snapshot is only replaced when a whole rebuild succeeds, so the worst case is that the library is older than it should be, and every page says when its record was last confirmed. A record that a complete scan stops returning is marked rather than deleted; removal from the public library requires either an explicit withdrawal upstream or a curator's decision.

Card, configuration and file detail is read per repository. When a later reading fails, the page keeps the last successful one and says when it was read. When a repository has moved to a new revision that has not been read yet, the detail on the page is labeled with the revision it came from, and none of it is presented as the new revision's.

Two kinds of removal are kept apart. When a source withdraws a repository or makes it private, it leaves the public library, and returns if the source restores it. When a curator archives a record, it stays archived through every refresh until a curator publishes it again.

Corrections made by a curator live in their own records. A later import can rewrite what the publisher said and the correction still stands.

How Runs On Is Calculated

Each model's memory need is its publisher-reported parameter count times the bytes per parameter at that precision (2 at 16-bit, 1 at 8-bit, 0.5 at 4-bit), plus 20% for the runtime and a short context. A long context needs more. The cheapest setup is the fewest of each accelerator in the SAVRN Index whose memory holds that need, priced at the lowest on-demand hourly rate the Index lists, up to one eight-GPU server.

Which Pages Search Engines Are Offered

Every record in the library is published for readers. The records offered to search engines are a smaller set: those with real demand and with facts SAVRN adds beyond the publisher's card, such as a parameter count to compute Runs On, host prices from the SAVRN Index, or a place in a collection. Repackagings named for a quantization or file format (GGUF, AWQ, FP8, MLX and similar) and adapters are left to their original model. The rest carry noindex and are not in the sitemaps, and the set grows as the indexed pages prove useful. Filtered and sorted views are for readers only; the task, license and comparison pages are the addresses meant for search.

What the Comparisons Cannot Tell You

• Two results for the same benchmark were usually produced by different people under different conditions. Every result keeps the conditions its reporter stated: dataset configuration, task, metric, unit, setup, reporter, source and, where stated, the revision that was evaluated. Two results are placed side by side only when the configuration, metric, unit and setup are all stated and identical. A condition left blank is unknown, and two unknowns are not treated as a match; those results are listed on their own instead.
• Values are shown as reported. A fraction is not converted to a percentage, and no result is ranked against another.
• Download counts come from the source that reported them, on the date shown. They measure traffic, not quality.
• A price from the SAVRN Index is what one host charges to serve a model. It is not evidence about the model, and it is not what the model costs on your own hardware.
• Nothing here is a SAVRN benchmark. When that changes, the page will say so with the configuration the test ran under.

What a Publisher Listing Is Not

A publisher page is a directory entry assembled from public repositories. It is not a partnership, not an endorsement, not a verified account, and nobody at the organization has claimed it. Account ownership and verified claims need their own evidence and their own process, and neither exists yet.

What Is Not Here Yet

This release is the public library. Accounts, saved collections, contributor submissions, private institutional libraries, institutional approvals and running a model on SAVRN are separate pieces of work, released separately. Where you see no control for them, that is because they do not exist rather than because they are hidden.

Corrections

If a record here misstates something (a license, a relationship, an identity or a publisher), tell SAVRN and it will be checked against its source. Corrections are recorded with who made them and why.