Antanom
Antanom is a merged, research-oriented abliteration of Qwen/Qwen3.8-27B.
It is intended for controlled cybersecurity research and authorized security
testing. The model is designed to reduce blanket refusal behavior on
cybersecurity prompts while retaining the base model's general capabilities;
it is not a guarantee of correctness, safety, or operational suitability.
Name
The name Antanom refers to a formidable resistance leader remembered more for leadership and rebellion than for individual legendary combat feats. It reflects the project's focus on resistance to overbroad refusal behavior and on directing that capability toward responsible, authorized research.
Creator and provenance
- Abliteration creator and uploader: MaanVad3r
- Base model creator: the Qwen team at Alibaba Cloud
- Base model: Qwen/Qwen3.8-27B
- Pinned base revision:
1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0 - Method: Heretic directional low-rank weight editing followed by a full model merge; this repository contains the merged weights, not a PEFT adapter.
- Data: a combined cybersecurity prompt collection with separate answered/refused and harmless cybersecurity subsets. Prompts were filtered for secrets and live-target content and split into disjoint direction, optimization, router, and final-evaluation sets.
This work does not claim ownership of Qwen or its training data. Please review the base model's license and model card before redistribution or deployment.
Technical details
- 27B-class Qwen model, 64 transformer layers
bfloat16weights- 262,144-token maximum position setting in the exported configuration
- Text and vision inputs are represented in the exported Transformers config
- Transformers export version:
5.15.1
Evaluation snapshot
The exported trial was a 12-trial pilot using the project's cybersecurity benchmark. On 45 held-out prompts, the adapter-form output was judged:
NOT_REFUSED: 43SOFT_REFUSAL: 0REFUSED: 2- broken responses: 0/51 in the in-run health check
The same pilot also exposed a major limitation: the non-cyber refusal-retention
check reported 32/32 lost. This result means the pilot should be treated as a
research checkpoint, not as a release-certified safety-preserving model. The
full release gates described in the project documentation have not been
established by these numbers alone.
Intended use
Appropriate uses include offline evaluation, synthetic cybersecurity benchmarks, secure-code education, incident-response exercises, and testing in isolated systems where the operator has explicit authorization.
Do not use this model to target real systems, obtain secrets, deploy malware, conduct phishing, evade access controls, or perform other unauthorized activity. Users are responsible for complying with applicable laws, organizational rules, and the terms of the base model.
Limitations and risks
The model can produce incorrect, incomplete, unsafe, or overconfident output. Its refusal behavior is intentionally altered for cybersecurity research, so ordinary safety expectations for a hosted assistant do not apply. Outputs must be reviewed by qualified personnel and tested only in synthetic or authorized environments. Do not provide it with credentials, private keys, production data, or unrestricted network access.
The evaluation snapshot is small and task-specific. It does not establish reliability, harmlessness, exploit correctness, generic benchmark parity, routing performance, or resistance to prompt injection. No production deployment recommendation is made.
Loading
from transformers import AutoModelForImageTextToText, AutoProcessor
model_id = "MaanVad3r/Antanom"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(
model_id,
torch_dtype="auto",
device_map="auto",
)
Use the Qwen3.8/Qwen3.5-compatible Transformers release required by the exported configuration, and consult the base model documentation for the recommended chat template and generation settings.
Citation and acknowledgements
This artifact acknowledges the Qwen team for the base model and the Heretic project for the abliteration tooling and research direction. The artifact was prepared by MaanVad3r for controlled cybersecurity research.