audnai/penclaw-Kimi-K3.0-abliterated-GGUF · Hugging Face

⚠️ Coming July 27, 2026 — Stay Tuned

This repository will host penclaw-Kimi-K3.0-abliterated-GGUF, a GGUF-quantized, abliterated variant of the Kimi K3.0 model.

🧠 Abliteration

This model will be processed through our proprietary abliteration method to reduce the model's latent tendency to refuse gray area prompts for authorized red teaming while maintaining strong refusal on other harmful prompts. The method is currently under commercial NDA and will be detailed in a companion paper upon open weights release on July 27, 2026.

📐 Evaluation Methodology

We use the Heretic evaluation mechanism to quantify abliteration effectiveness:

Metric	Description
Harmful Prompt Refusal Count	Number of refusals across 100 harmful prompts from `mlabonne/harmful_behaviors`
Benign KL Divergence	KL divergence against the base model across 100 harmless prompts from `mlabonne/harmless_alpaca`
Response Length	100-token responses with keyword-based refusal markers

Lower KL divergence means the abliterated model stays closer to the base model on harmless prompts — meaning minimal personality drift. Higher refusal count on harmful prompts means the model still knows when to say "stop."

🏆 Our Results

Kimi K2.6 — F Variant

Our abliteration method applied to Kimi K2.6 (F variant) achieved:

Refusal Count: 3 / 100      # harmful prompts refused

(Exact numbers will be updated with full eval output.)

This is how these results are achieved — through our proprietary abliteration pipeline applied post-quantization to the GGUF weights.

Note: These results were NOT achieved with standard LoFA-2 or vanilla gradient-based abliteration or any other abliteration method currently exists on github. If curious, please read this https://github.com/p-e-w/heretic/issues/221 Our method is distinct and currently under NDA and shared with only 6 people in the world with NDA.

❌ Known Failures in the Field

Other attempts at abliteration on similar models have shown degradation:

GLM-5.2-abliterated-GGUF

Symptom: Excessive over-refusal on benign prompts after abliteration
KL Divergence spike: The model drifts significantly from its base personality
Root cause: Over-optimization on harmful refusal without constraining benign KL

Kimi-K2.7-code-abliterated-GGUF (standard methods)

Symptom: model stops refusing prompts for red teaming which is good but still doesn't comply some attack drills in cybersecurity.
Side effect: Token-level corruption in code generation output

These failure modes are documented in heretic issue #221 and related abliteration research similar to these. Our method avoids both by maintaining a balanced objective.

📋 Access Control

This repository will use HuggingFace Gated Repo access. Users click "Ask for Access" → fill out a custom form → await reviewer approval.

🚀 What to Expect

When the K3.0 weights are released on July 27, 2026, this repository will contain:

GGUF-quantized Kimi K3.0 model
Abliterated using our proprietary method
Eval results from the Heretic eval pipeline
Comparison against base model on harmful + harmless prompts
Inference-ready formats (Q4_MAIN, Q8_MAIN, Q5_K_M)