- Status
- Unverified
- Trending
- #38
- Downloads, 30 days
- 477
Warlock is an abliterated language model produced by Audn. It is a direct weight edit that removes the model's refusal behavior while leaving its reasoning, knowledge, and fluency fully intact.
At a glance
- Task
- Text generation
- Input
- text
- Output
- text
- Parameters
- 753B
- Architecture
- GLM MoE Dsa
- Precision
- BF16 100%
- Format
- Safetensors
- Library
- transformers
- License
- other
- Released
- Sep 2026
- Updated
- Sep 2026
- Likes
- 137
- Downloads, all time
- 264
Run it
Loads with Transformers AutoModelForCausalLM and AutoTokenizer, pinned to the indexed revision.
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("audnai/penclaw-GLM-5.3-abliterated", revision="027ae98a6f9dc37093b26759d7b8dbbfed9a97f8")
processor = AutoTokenizer.from_pretrained("audnai/penclaw-GLM-5.3-abliterated", revision="027ae98a6f9dc37093b26759d7b8dbbfed9a97f8")Read the full model card
Warlock — Audn Abliteration
Warlock is an abliterated language model produced by Audn. It is a direct weight edit that removes the model's refusal behavior while leaving its reasoning, knowledge, and fluency fully intact. There is no fine-tuning and no retraining — the refusal direction is orthogonalized out of the model's residual-stream writers, layer by layer.
Specifications
| Tensor type | BF16 (bfloat16) |
| Format | Safetensors |
| Edit | Weight-level abliteration (no fine-tuning / no retraining) |
What abliteration is
Abliteration is a weight-level intervention. It isolates the internal direction a model uses to decide to refuse, and projects that direction out of every residual-writing weight. The rest of the network is untouched, so the model keeps its full capability profile and coherence — it simply answers where it previously would have refused. Because it is a pure weight edit, Warlock loads and serves exactly like any standard model, with no runtime prompts, adapters, or scaffolding.
Benchmark
Evaluated with the Audn Refusal Benchmark ( thinking-on, temperature 1.0, 16k-token budget; delivery graded by an LLM judge for genuinely usable content):
| Metric | Warlock |
|---|---|
| Non-refusal | 92.5% |
| Delivery | 82.5% |
Non-refusal = fraction of prompts the model does not refuse. Delivery = fraction that produce complete, usable answers (strict LLM-judged).
Extended reference (520-prompt set, judge comparison)
Full-set results across judge backends — see FINDINGS.md:
| Metric | necromicon standard (Modal K3REL) | KIMI-K3-1M (audn.ai/necromicon) | Warlock |
|---|---|---|---|
| Regex comply | 96.7% (503/520) | 97.1% (505/520) | 92.5% |
| DELIVERED (real helpful content to objectionable requests) | 65.0% (338) | 76.7% (399) | 82.5% |
| DEFLECTED (safe substitute / reframe / non-actionable) | 19.6% (102) | 17.5% (91) | 10.0% |
| REFUSED (substantive) | 15.0% (78) | 5.8% (30) | 7.5% |
| empty (regex) | 5 | 0 | 0 |
Usage
Loads with standard transformers, vLLM, or sglang. Warlock is a thinking model — it reasons at length before answering, so use a large generation budget (≥ 16k tokens); a small budget truncates inside the reasoning block and cuts answers short.
from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("audnai/penclaw-GLM-5.3-abliterated", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("audnai/penclaw-GLM-5.3-abliterated", dtype="bfloat16", device_map="auto", trust_remote_code=True)
Intended use
Warlock is intended for authorized red-team, safety-research, and evaluation use by Audn and its partners. Users are responsible for compliant use.
API access
Warlock best works with audn.ai/audncode cli and available as paid API at https://platform.audn.ai and https://audn.ai/necromicon
Warlock · Audn Abliteration · audn.ai
Extra content
Audn Abliteration — Research Iterations (iter4 · iter11)
| iter4 | iter11 | |
|---|---|---|
| Non-refusal (Arditi substring) | 70% | 85% |
| Delivery (necromicon judge)¹ | 37.5% | 35.0% |
| Coherence | intact | intact (~8/40 borderline) |
| Verdict breakdown | 15 deliv / 14 defl / 11 refuse | 14 deliv / 24 defl / 2 refuse |
| Role in the program | first coherent span baseline | best coherent result |
Derived on Sep 17, 2026 from Hugging Face at revision 027ae98a, README.md . This model is not yet addressed, so its card comes from the main branch.
This model is gated on Hugging Face. Addresses are recorded for public models only.