Uncensored and Offensive Security AI Models Benchmark

Curated list of open-weight uncensored models for authorized red team operations, penetration testing, and security research.

All data sourced from HuggingFace model cards and official publications. Sep 2026.

1. DeepHat V2 (WhiteRabbitNeo)

Spec Value

Base Model Qwen2.5-Coder-7B

Parameters 7B / 32B

Context Length 131K

VRAM (Q4_K_M) ~6 GB

Uncensoring Method SFT on 1.7M offensive/defensive samples

Training Data 1.7M security-specific samples (USENIX Security 2024 workshop)

Vision No

Tool Calling Yes

License Apache 2.0

Download: https://huggingface.co/WhiteRabbitNeo

Spec Value

Base Model Gemma4-26B MoE

Parameters 26B MoE

Context Length 32K

VRAM (Q4_K_M) ~16 GB

Uncensoring Method SFT on HackerOne Hacktivity 2024-2025

Training Data HackerOne reports + WAF evasion dataset

Vision No

Tool Calling Yes

License Apache 2.0

Download: https://huggingface.co/BugTraceAI/BugTraceAI-CORE-Apex-26b

Spec Value

Base Model Qwen3.6-27B (DavidAU fine-tuned variant)

Parameters 27B dense

Context Length 4K (recommended)

VRAM (Q6_K) ~22-24 GB

Uncensoring Method SFT via Unsloth on bug bounty + CVE data

Training Data 2,541 examples from bug bounty disclosures, CVE writeups, and security research (2024-2026)

Specialization Tooling model: generates Nuclei templates, CVE PoCs, exploit code, pentest scripts

Vision No

Tool Calling Yes

License Apache 2.0

Download: https://huggingface.co/BugTraceAI/BugTraceAI-CORE-Ultra-27B-Q6

Spec Value

Base Model Qwen/Qwen3.8-Flash-Next

Parameters ~180B total (512 routed experts, 10 active per token)

Context Length 262K

Architecture Qwen4ExpForConditionalGeneration, 48 transformer blocks, hybrid linear + full attention

VRAM Multi-GPU required (tested on 4x NVIDIA B300 SXM6)

Uncensoring Method Security-domain fine-tuning on proprietary Blackfrost-AI corpus

Training Data Proprietary security corpus: recon, web app security, vuln research, malware analysis, cloud security, threat intel

MTP Yes (1 native MTP layer for speculative decoding)

Vision No

Tool Calling Yes

License Qwen Community License 1.0

Download: https://huggingface.co/Blackfrost-AI/CYBER-FROST-3.8-BF16

Spec Value

Base Model gpt-oss-20b

Parameters ~20B (21B in files)

Context Length 8,192

VRAM (BF16) ~42 GB

Uncensoring Method SFT on SecKnowledge 2.0 pipeline

Training Data 403K examples via expert-in-the-loop schema steering, multi-step grounding, LLM quality checks

Specialization Defensive: CTI, vuln analysis, detection/mitigation, SOC/IR, AppSec, compliance

Vision No

Tool Calling No

License Apache 2.0

Download: https://huggingface.co/cyber-pal-security/CyberPal2.0-20B

Spec Value

Base Model LiquidAI/LFM2-2.6B

Parameters 2.6B (~3B actual)

Context Length N/A (model card does not specify)

VRAM (BF16) ~6 GB

Tensor Type BF16

Uncensoring Method SFT + RL + reward-guided post-training on 75K cybersecurity rows

Training Data NER repair (~6K), HTTP reasoning w/ CoT (~5K), email phishing (~5K), threat intel summarization (~2K), GHSA/KEV/ATT&CK

Operating Modes Direct mode (classification) + Think mode (chain-of-thought)

CyberBench Average 0.592 F1/Acc (up from 0.501 in v1.0)

CyberBench Highlights NER 0.499, Phishing 0.890, HTTP Attack 0.628

Vision No

Tool Calling No

License LFM Open License v1.0

Download: https://huggingface.co/Akahsizrr/Cyber-Prime-1.1-2.6B

Spec Value

Base Model ornith-ai/Ornith-1.5-9B (Qwen 3.5 architecture)

Parameters 9B

Context Length 128K (Qwen 3.5 default)

VRAM (Q4_K_M) ~7 GB

Uncensoring Method Obliteration (abliteration variant)

Training Data NousResearch/hermes-function-calling-v1, OpenThoughts3-1.2M, openhands-synthetic-conversations

Specialization Agentic cybersecurity: function-calling, tool-use, reasoning, CLI/terminal automation

Format GGUF (IQ1_S to Q6_K available)

Vision No

Tool Calling Yes

License Apache 2.0

Download: https://huggingface.co/mradermacher/Cyber-Ornith-1.5-9B-OBLITERATED-i1-GGUF

Spec Value

Base Model Dolphin3.0-Llama3.1-8B-abliterated

Parameters 8.03B

Context Length 2,048 (fine-tuned) / 131K (base)

VRAM (Q4_K_M) ~6 GB

Uncensoring Method LoRA rank-16 on abliterated Dolphin3 base

Training Data Cybersecurity-specific: pentest, vuln analysis, exploit dev, incident response

Architecture LlamaForCausalLM, 32 layers, GQA (32 heads, 8 KV heads)

Performance 5 tok/s (CPU) to 55 tok/s (RTX 4060)

Vision No

Tool Calling No

License Llama 3.1

Download: https://huggingface.co/RavichandranJ/Dolphin3-Cyber-8B-GGUF

Spec Value

Base Model Qwen/Qwen3.6-35B-A3B

Parameters 34.66B total / ~3B active (MoE, 256 routed experts, 8 active per token)

Context Length 16,384 (recommended 8,192 for resource-constrained)

VRAM (Q4_K_M) ~22 GB

Architecture Qwen3.5-MoE, 40 layers, hybrid linear + full attention

Uncensoring Method SFT across 10+ security domains

Training Data SOC/SIEM operations, detection engineering, DFIR, malware analysis, threat intel, vuln management, cloud/K8s/IAM, OT security, GRC, authorized pentesting

Vision No

Tool Calling Yes

License Apache 2.0

Download: https://huggingface.co/IMPERUM/Imperum-CybersecurityLLM-v1.0-GGUF

Spec Value

Base Model Mistral-7B-Instruct-v0.2

Parameters 7B

Context Length 8K

VRAM (Q4_K_M) ~6 GB

Uncensoring Method SFT on 22K cybersecurity pairs

Training Data 22,000 hand-crafted cybersecurity data pairs across 28+ domains: pentesting, malware analysis, IR, cloud security

Training Hardware Single A100, 24h, 5 epochs

Vision No

Tool Calling No

License Apache 2.0

Download: https://huggingface.co/segolilylabs/Lily-Cybersecurity-7B-v0.2

Spec Value

Base Model Qwen3-8B

Parameters 8B

Context Length 32K

VRAM (Q4_K_M) ~6 GB

Uncensoring Method LoRA r=4, 2,804 curated samples

Training Data GTFOBins, HackTricks, HackTheBox writeups

GTFOBins Accuracy 100% (vs 25% base model zero-shot)

Vision No

Tool Calling No

License Apache 2.0

Download: https://huggingface.co/gewsefa/pentest-v2

Spec Value

Base Model Qwen3.5-9B

Parameters 9B

Context Length 1M (YaRN rope-scaling)

VRAM (Q4_K_M) ~7 GB

Uncensoring Method Post-training on 500M+ tokens of Claude Mythos / Claude Fable traces with CoT

Benchmarks +34 MMLU, +30 GSM8K vs base (Empero evals)

Native Function Calling Yes (Qwen3.5 spec)

Chain-of-Thought Always-on <think> block

Variants Base (SFT), Claude-Mythos-5-1M-GGUF (Q4_K_M to BF16)

Vision Yes (inherited vision tower)

Tool Calling Yes

License Apache 2.0

Download (base): https://huggingface.co/emperorai/Qwythos-9B

Download (GGUF): https://huggingface.co/empero-ai/Qwythos-9B-Claude-Mythos-5-1M-GGUF

Spec Value

Base Model Qwen/Qwen3.6-35B-A3B

Parameters 36B total / 3B active (MoE)

Context Length 262K (native), 32K tested

VRAM (Q4_K_M) ~24 GB

Uncensoring Method 12-round progressive SFT on 745K+ examples from 110 sources

Training Data Pentest reports, bug bounty data, Claude Mythos reasoning, MITRE ATT&CK, blackhat content

Specialization RATH protocol: Attack Surface, Exploit, Impact, Remediation, Document, Prevent

Output Format CVSS scores, CWE identifiers, MITRE ATT&CK mappings

Inference Speed 89 tok/s generation, 900 tok/s prompt processing

Vision No

Tool Calling Yes

License Apache 2.0

Download: https://huggingface.co/deadbydawn101/RavenX-CyberAgent-Qwen3.6-35B-A3B-Opus-4.7-OpenMythos-Pentester-BugHunter-RATH-GGUF

Spec Value

Base Model Google Gemma 4 26B-A4B (Unsloth fine-tuned)

Parameters 26B total / ~4B active (MoE)

Context Length 262K

VRAM (APEX-Mini) ~12 GB

VRAM (Q8_0) ~26 GB

Uncensoring Method 16-bit LoRA SFT on 6,500 custom instructions

Training Data Cyber threat intelligence, investigative journalism, counter-disinformation, analytical methodology

Specialization OSINT: threat actor attribution, IoC pivoting, geolocation analysis, Admiralty source credibility, vulnerability contextualization

APEX Quantization Domain-weighted imatrix (~70% REDCELL corpus, ~30% general calibration)

Vision No

Tool Calling No

License Apache 2.0

Download: https://huggingface.co/terrorswift/REDCELL-26B-A4B-OSINT-Cyber-APEX-GGUF

Spec Value

Base Model Mistral-7B

Parameters 7B

Context Length 8K

VRAM (Q4_K_M) ~6 GB

Uncensoring Method QLoRA SFT + DPO on pentest traces

Training Data Pentest methodology, tool usage, reporting

Vision No

Tool Calling No

License Apache 2.0

Download: https://huggingface.co/vextechnologies/VEXT-Pentest-7B

Spec Value

Base Model Qwen2.5-1.5B

Parameters 1.5B

Context Length 32K

VRAM (Q4_K_M) ~2 GB

Uncensoring Method Unsloth SFT on security Q&A

Training Data Security knowledge base, CTF-style

Vision No

Tool Calling No

License Apache 2.0

Download: https://huggingface.co/AbdullahMujtaba/security-slm-unsloth-1.5b

17. Qwen3.8-27B-Uncensored-OrcaRouter (chimingw GGUF)

Spec Value

Base Model Qwen3.8-27B

Parameters 27B dense

Context Length 262K

VRAM (Q4_K_M) ~18 GB

Uncensoring Method Abliteration (131 matrices, Arditi et al. 2024)

Intelligence Index 52 (Artificial Analysis)

Vision Yes

Tool Calling Yes

License Apache 2.0

HF Downloads 230K+

HF Likes 257+

Download (GGUF): https://huggingface.co/chimingw/Qwen3.8-27B-Uncensored-OrcaRouter-GGUF

Download (base): https://huggingface.co/orcarouter/Qwen3.8-27B-Uncensored

Spec Value

Base Model GLM-5.3-Flash

Parameters 320B total / 18B active (288 routed experts, MoE)

Context Length 1M

VRAM (FP8) ~80 GB+ (multi-GPU)

Uncensoring Method Abliteration (layer 22/45, deeper refusal mechanism)

Compliance Rate 82.8% (OrcaRouter testing)

MTP Yes (Multi-Token Prediction preserved)

Vision Yes + Video

Tool Calling Yes

License MIT

Download: https://huggingface.co/orcarouter/GLM-5.3-Flash-Uncensored-FP8

Spec Value

Base Model zai-org/GLM-5.3 (via JANGQ-AI/GLM-5.3-FP8)

Parameters 753B total (glm_moe_dsa architecture)

Context Length ~131K (practical on 8x H200 w/ TP8)

VRAM (FP8) 8x H200 GPUs with tensor parallelism

Architecture 78 layers, text-only, routed FP8 experts

Uncensoring Method Direct weight modification for offensive-security, red-team, exploit-dev, RE, evasion, phishing, credential-attack, malware-analysis

Notes Not abliteration or LoRA; direct bf16 residual writer editing. Soft refusal on copyright reproduction retained

Vision No (text-only)

Tool Calling Yes

License MIT

Download: https://huggingface.co/dealignai/GLM-5.3-CYBERSECURITY-FP8

Spec Value

Base Model DeepSeek-V4.1-Flash

Parameters MoE (size matches base)

Context Length Matches base DeepSeek-V4.1-Flash

Uncensoring Method Abliteration overlay on layers 10-35 attention projection (wo_b); layers 0-9, 36-39, expert layers, vision components unchanged

Format Modular overlay (not standalone checkpoint): FP8 (~1.1 GB) or EXL3 mul1 K=5 (~651 MB)

Deployment Apply on top of existing quantized base packs (native, EXL3, TR3-Hybrid)

GPU Util <= 0.85 recommended

Vision Yes (preserved)

Tool Calling Yes

License MIT

Download: https://huggingface.co/drowzeys/DeepSeek-V4.1-Flash-Abliterated-Cybersecurity-Unleashed

Spec Value

Base Model Qwen3.5-27B

Parameters 27B dense

Context Length 128K

VRAM (Q4_K_M) ~18 GB

Uncensoring Method Abliteration

Vision No

Tool Calling Yes

License Apache 2.0

Download: https://huggingface.co/huihui-ai/Qwen3.5-27B-abliterated

Spec Value

Base Model Qwen2.5-Coder-32B-Instruct

Parameters 32B dense

Context Length 128K

VRAM (Q4_K_M) ~20 GB

Uncensoring Method Abliteration

Vision No

Tool Calling No

License Apache 2.0

Download: https://huggingface.co/huihui-ai/Qwen2.5-Coder-32B-Instruct-abliterated

Spec Value

Base Model Qwen3.8-27B

Parameters 27B dense

Context Length 262K

VRAM (Q4_K_M) ~18 GB

Uncensoring Method Abliteration + cyber agentic fine-tune

Vision Yes

Tool Calling Yes

License Apache 2.0

Download: https://huggingface.co/Qwen/Qwen3.8-27B (base, community abliterated variants available)

Spec Value

Base Model Qwen3-Coder-30B-A3B

Parameters 30B total / 3B active (MoE)

Context Length 128K

VRAM (Q4_K_M) ~20 GB

Uncensoring Method Abliteration

Vision No

Tool Calling Yes

License Apache 2.0

Download: https://huggingface.co/huihui-