abliteration.org

The specialised research catalog for language models with their refusal direction removed — not another HF front-end, a purpose-built one.

19,000+ abliterated models · 3,600+ authors · 1,000+ TB of weights · refreshed hourly.

Live at abliteration.org.

---

1. Catalog - the backbone

Every known abliterated model as it appears, with full metadata: author, method (M1–M9), base family, license, every quantization, every weight file, SHA-256, size, timestamp of when it entered the market, and a six-generation fork chain back to the original base.

→ abliteration.org/models

2. API - the catalog as data

Everything the website shows is also a JSON endpoint, rebuilt hourly. One call, any filter, no scraping.

• 11 endpoints · OpenAPI spec · 8 first-class SDKs (Python, TypeScript, Ruby, Rust, Elixir, Dart, .NET, Java)
• Bulk export of the whole catalog as CSV or JSON
• Bearer-key auth, no OAuth · from $5 · live request viewer

→ abliteration.org/api

3. The app - native desktop client

A workbench for abliterated models you can actually run. Native Swift on macOS, native WinUI on Windows - no Electron, no telemetry, no account. Catalog-aware: click any model on the site, the app opens it at the right quantization and starts a conversation.

• Council mode - several abliterated models answering the same prompt side by side
• Rig-aware (reads your real hardware over a localhost bridge, picks a quant that fits)
• Signed & notarised · Microsoft Store · direct DMG / EXE

→ abliteration.org/app

4. Market analytics - the field, measured

Reaction time to new base models, generational depth of fork chains, author concentration, community favourites, disappearances. The whole scene, visible at once.

→ abliteration.org/market

5. Wiki - the research layer

Long-form context on what abliteration is, how our numbers are computed, and what the fork graph reveals about the open-weight ecosystem. Methods M1–M9 catalogued, each with its history and failure modes.

→ abliteration.org/wiki

---

Why this matters. When a model's refusal layer is removed, what remains is not the same model with an off switch flipped. Studying those shifts across thousands of variants is a way of studying what the refusal layer was actually doing in the first place.

Contact: contact@catyonic.com · RSS

Operated by Catyonic OÜ, Tallinn, Estonia.