Vectors / Methods

V8: Experimental or undocumented

The author does not describe the method clearly, or uses something new that fits none of V1 to V7.

Original source: Varies per repo

Repos

4

Vectors

21

SAE-based cutting-edge

Steering directions refined with a sparse autoencoder (BatchTopK SAE), a current research technique.

amarck/Qwen3.5-35B-A3B-abliterated-GGUF

V8 cutting-edge Qwen · apache-2.0 · latest commit 2026-03-07

Source cited: Arditi et al. 2024 (abliteration of the model weights); custom SAE pipeline (BatchTopK SAE) for the control vector

1 vector 1 set · 298 KB
229 repo downloads

Undocumented / experimental

gghfez/DeepSeek-V4-Flash-control-vectors

V8 DeepSeek · cc-by-nc-4.0 · latest commit 2026-07-21

Source cited: jukofyork creative-writing-control-vectors-v3.0 (cross-covariance eigendecomposition)

18 vectors 6 sets · 11.2 MB
260 repo downloads

andyoneal/Plainspeak-Dial-Gemma-4-26B-A4B

V8 Gemma · apache-2.0 · latest commit 2026-10-23

Source cited: llama.cpp --control-vector GGUF; no paper or library cited in the visible README

1 vector 1 set · 320 KB
76 repo downloads

26B-Suite/Plainspeak-Dial-Gemma-4-26B-A4B-BakeVector

V8 Gemma · gemma · latest commit 2026-08-15

Source cited: none mentioned; used with llama.cpp --control-vector

1 vector 1 set · 320 KB
64 repo downloads

Catalog is the map. Apps are the tools.

Run models on your own machine, not in the cloud.

Every model page has an "Open in Abliteration" button that hands the model directly to the first-party desktop client, at the quantization your rig can actually run. No API keys, no subscription, no prompt leakage.

Get the app Browse the catalog