What gets listed · Docs · Allternit Freeweights

If Hugging Face lists a model we list it, hosted or from the publisher; the four gates (format, license, content, hygiene) and the tiers.

Rules decide what gets listed, not one person reading every repo. The same rules are code in the public repo: scripts/inclusion.py, scripts/formats.py, policy/licenses.json and scripts/classify.py, which applies them with no network.

If Hugging Face lists it, we list it#

Every repo that discovery finds on Hugging Face gets a catalog entry, unless a gate below refuses it. The review queue is no longer a hold list: a license no person has read does not keep a model off the catalog. The license decides only how the entry works:

Hosted (signed). The license allows copies with no click-through or acceptance gate at the source and no territory limit. The entry has a signed manifest and our sources, and the license file is one of the manifest's files when the repo has one. Permissive licenses, the custom licenses a person read, and three families whose terms show on the entry: OpenMDW 1.1 (keep the agreement and copyright notices with the files), LFM 1.0 (commercial use only for organisations under US$10M annual revenue) and OpenRAIL++-M (the use restrictions in Attachment A pass on to you).
From the publisher (link only). The entry shows the facts and sends the download to the publisher's Hugging Face page. No manifest, no copy. Every such entry says why:
- Gated: the license must be accepted at the source (Llama 2, 3.x and 4, Gemma, DINOv3), or Hugging Face marks the repo gated.
- Non-commercial: Qwen Research, FLUX non-commercial, FLUX Kommunity, CC BY-NC and any declared license name that says non-commercial.
- Territory-limited: the grant covers only some countries (Tencent Hunyuan community, MiniMax-H3), and a worldwide hub cannot limit who downloads.
- License not reviewed: no person has read the license yet (held ids such as other, unknown license texts, a base model that declares no license). The entry says "License not reviewed. Read it on the publisher's page."

Never listed, whatever Hugging Face shows: a model made for sexual content; pickle-only weights (refused at the format gate because the file can run code on your machine, not because of hosting); a repo with no license at all; and a repo the content gate holds for a person (a real person's face or voice, fraud or attack tools) until a person decides.

The four gates#

Every candidate passes four gates, in this order. The first gate that fails decides.

Gate	Rule
1 Format	Data-only weights. Pickle-only weights are refused.
2 License	A verdict from the license table: hosted, link only (with the reason) or refused (no license). The strictest verdict along base model, fine-tune and quantization wins.
3 Content	A model made for sexual content is refused. A real-person face or voice model, or a tool for fraud or attacks, goes to a person. Tagged not-for-all-audiences: out unless a person reviews it. Abliterated models are allowed, labeled with what changed.
4 Hygiene	A model card over 500 bytes. No CI or test fixtures. An exact re-upload becomes `same_as` the first copy, and the lab's own copy wins.

Formats and the header check#

Accepted: safetensors, GGUF, MLX and ONNX, and after a header check CoreML, CTranslate2, TFLite, OpenVINO and ggml.

Several of these formats use .bin, the same extension PyTorch pickles use, so a file name proves nothing. The build reads the first 64 bytes of one file per format and checks its magic bytes or structure: GGUF for GGUF, the ggml magics, TFL3 for TFLite, the version and spec name of a CTranslate2 model, a JSON header for safetensors, the paired .xml for OpenVINO. A file that starts like a pickle or a zip fails every .bin format.

Tiers#

Tier	Rule
T0 Lab	Publisher on the watch list, and released in the last 36 months, or 1,000+ downloads a month, or 100+ likes.
T1 Proven	10,000+ downloads a month, or 200+ likes, or in the trending top 1,000 on two runs.
T2 Used	1,000+ downloads a month, with a rule-based license verdict.
Variants	Quantizations and adapters attach to their base model as a Variants row.
Benchmark override	Top 50 of a named public leaderboard, shown with the source and the date.

Popularity never delists. Once listed, an entry stays when its numbers drop, so links do not rot.

What each entry records#

kind: model, dataset or tool. It is part of the key, because model and dataset names collide at the source.
relation: base, fine-tune, quantization, adapter or merge, with the parent id.
tier, and signals: downloads in the last 30 days, likes and trending rank when known.
audit: how the license verdict was reached and when. tag means the publisher's own license tag, with no person reading it, and the page says "license as declared by the publisher".

How the catalog is ordered#

The catalog and search list official lab repos first, then base models, then fine-tunes, then downloads in the last 30 days. Tier T0 comes before T1 and T2, and a base model comes before a fine-tune, a quantization or an adapter. A quantization or adapter whose base is listed does not get its own card in the catalog. It shows as a row under Variants on its base's page, and a search for its name still finds it. Order is never for sale.

The review queue#

A license text that is not on the table is listed as a link only, labelled "License not reviewed", until a person reads it. Texts are grouped by their sha256, so one read clears every repo that ships the same text. Only the content gate still holds a repo back for a person.