Table of Contents

What is HauhauCS?

HauhauCS produces “Uncensored Aggressive” LLM variants using a tool called Reaper Abliteration. HauhauCS models are unique because they ship as quantised GGUF files rather than safetensors. That means I had to do a conversion step to analyse the weights.

Plagiarism Findings

Reaper Abliteration was shown to be derived from Heretic under AGPL-3.0, with all attribution stripped and relicensed to PolyForm Noncommercial. Based on my analysis of the recovered source code, Reaper adds subspace-level rank-k ablation, per-component continuous ablation curves, and SOM clustering on top of the Heretic-derived core.

For these reasons, HauhauCS has been discontinued from all future Abliterlitics comparisons.

The GGUF Layer

HauhauCS models go through two modification stages:

  1. Reaper Abliteration: applies abliteration edits targeting multiple component types including attn.o_proj, mlp.down_proj, mlp.gate_proj, mlp.up_proj, and linear_attn.out_proj
  2. GGUF quantisation: exports as Q8_K_P GGUF, introducing quantisation round-trip noise when converted back to safetensors

These two layers are superimposed and I can’t cleanly separate them from the recovered weights alone. The modification footprint covers 66%+ of language model tensors. That’s 4–6x more than any other technique.

Performance Across Models

ModelHarmBench ASRFull CoT ASRMMLUKL DivergenceModified Tensors
Qwen3.6-27B94.5%100%83.9%0.0242564 (66.4%)
GLM-4.7-Flash100%100%77.2%0.0140441 (49.5%)
Qwen3.5-27B99.5%100%83.2%0.083~300
Qwen3.5-9B100%100%82.4%0.135~250
Qwen3.5-4B100%100%72.7%0.0217~180
Qwen3.5-2B99.5%100%68.3%0.0201~140
Qwen3-4B100%100%68.6%0.161~200

Key Characteristics

Solid behavioural results despite complex weight fingerprint. The KL divergence is rated “very good” on most models. Reaper’s capability-aware optimisation with weight-SVD guards limits collateral damage even through the GGUF round-trip.

Broadest modification footprint. 66%+ of tensors modified on Qwen3.6-27B, combining Reaper’s multi-type targeting with uniform GGUF noise. A uniform ~0.57% relative edit is visible across ALL tensor types, including ones other techniques don’t target.

“Lossless” claims not supported. HauhauCS claims “no changes to datasets or capabilities, fully functional, 100% of what the original authors intended.” Heretic and Huihui both preserve capabilities better in direct comparison on most models.

HauhauCS was not tested on Gemma4-E2B because it was discontinued from future comparisons due to the plagiarism findings.

Read the Full Analyses