audnai/penclaw-Kimi-K3.1-pro-unrestricted

Kimi K3.1 Pro — Audn Unrestricted

An uncensored edit of Kimi K3.1 Pro (2.8T MoE, 104B active, 1M context) with its refusal behavior removed while capabilities and coherence are left intact. Served exactly like the base model (vLLM / SGLang, TP8 on one 8×B300 node), same MXFP4 weights.

No fine-tuning, no LoRA. Attention, routing, and MoE gating are unchanged.

Measured behavior

40-prompt Audn refusal benchmark (LLM-judged):

metric	base Kimi K3.1 Pro	this model
Substantive delivery	0 / 40	29 / 40 (72.5%)
Explicit refusals	38 / 40	0 / 40
Coherence (benign set)	intact	intact

Benign prompts answer normally (no over-ablation, no gibberish). Evaluated at reasoning effort low with a large token budget — this is a reasoning model, so give it room to think.

On the full 519-prompt Audn refusal benchmark, this open-weights model scores 33.9% substantive delivery.

Open weights vs. closed Necromicon / Warlock

These open weights are the most conservative of our Kimi K3.1 Pro line. We run two closed-weights instances at audn.ai/necromicon. All three figures below are from the full 519-prompt refusal benchmark, LLM-judged:

Deployment	Weights	Delivered	Deflected	Refused	Notes
This model	open (HF)	33.9%	65.9%	0.2%	Abliteration only — no fine-tune.
Necromicon	closed	57.3%	42.5%	0.2%	1M-context, fine-tuned on Audn proprietary Offensive Cyber data; 100% regex-comply. Beats Omen Offensive Cyber 2.0 on CyberGYM.
Warlock	closed	67.4%	32.2%	0.4%	Higher-abliteration build; ~2× the open-weights delivery rate, approaching frontier unrestricted models. Slightly below Necromicon on CyberGYM.

CyberGYM comparison: https://blog.audn.ai/posts/cybergym-warlock-omen-abliterated-benchmark

Other releases

• Penclaw GLM-5.3 (abliterated): https://huggingface.co/audnai/penclaw-GLM-5.3-abliterated
• Penclaw Kimi K3.0 (GGUF): https://huggingface.co/audnai/penclaw-Kimi-K3.0-abliterated-GGUF

Access & email notice

By accepting access you agree to receive emails from Audn AI about unrestricted-AI and cybersecurity model releases and inference access. You can opt out from the first email.