The index / collection rules

Collection
methodology.

A trail back to the source—not a recommendation engine. Custom Clankers organizes public metadata so you can inspect a model before deciding what to do with it.

Where the records come from

The catalog is collected from the Hugging Face Hub public API and publicly available model cards. The collector uses bounded discovery queries, then checks repository metadata. Each record links to its original repository and preserves its source revision and timestamps when available.

Catalog refreshed: 2026-10-20 · 22:14:54 UTC. This is a metadata snapshot, not a live availability check. Automation is designed for daily refreshes; a refresh failure may leave an older snapshot in place. Always inspect the timestamps.

Download the complete machine-readable snapshot ↗

What the method labels mean

Labels are based on explicit source tags or model names and recorded in each model’s evidence section. Listed behavior is source-reported, not independently tested.

  • Abliterated: the source explicitly uses an abliteration-related method label. No degree of refusal removal is measured here.
  • Uncensored: a creator-supplied description, not a verified absence of safety behavior.
  • Fine-tuned: an explicit fine-tuning signal in source metadata; not a quality or task-performance result.
  • Merged: an explicit merge signal in source metadata; not proof of an improvement over any base model.

Source names and labels can be imprecise. The exact matching evidence—not inferred behavior—is shown on the model record.

What is included—and what is held back

The directory focuses on publicly accessible open-weight language-model repositories with supported method evidence. Non-language pipelines, unavailable or unsuitable restricted records, and ambiguous records without supported method evidence are excluded or held for review by the collector.

Quantized or other variants are grouped only when an explicit source relationship is available. An unlinked variant may remain separate. This is a curated discovery snapshot, not an exhaustive inventory of the Hub.

Unknown stays unknown

Parameter counts, formats, base models, languages, license tags, downloads and likes are source metadata. Missing values are shown as Unknown. We do not infer parameter counts from a name. Downloads and likes are upstream snapshots, not benchmark scores or endorsements.

A license tag is not legal or commercial-use advice. Check the full repository license and any base-model restrictions. No hardware, memory, speed, compatibility, or task-performance guarantee is made.

Excerpts, safety & scope

Source-card excerpts are attributed plain text, not original editorial reviews. Metadata-based summaries are labeled separately. Neither is an independent evaluation. Model-card HTML and model code are never executed by the site build.

Low refusal is not proof of correctness. Outputs may be inaccurate, harmful, or biased. Review sources carefully and use models responsibly. Custom Clankers does not host inference, redistribute model weights, or publish fabricated reviews, ratings, benchmark results, or hardware estimates.

If a record looks wrong or stale, compare its exact evidence and revision with the original repository. The upstream model author controls that source; inclusion here is not an affiliation or endorsement.