JoasASantos/Offensive-Security-AI-Models

Uncensored LLMs for Offensive Security

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.

---

Security Fine-tuned Models

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

---

2. BugTraceAI-CORE-Apex (26B)

| 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

---

3. BugTraceAI-CORE-Ultra (27B)

| 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

---

4. CYBER-FROST-3.8 (Blackfrost-AI)

| 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

---

5. CyberPal 2.0 (20B)

| 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

---

6. Cyber-Prime 1.1 (2.6B)

| 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

---

7. Cyber-Ornith-1.5-9B (DuoNeural / mradermacher)

| 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

---

8. Dolphin3-Cyber-8B (RavichandranJ)

| 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

---

9. Imperum-CybersecurityLLM v1.0

| 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

---

10. Lily-Cybersecurity-7B v0.2 (Segolily Labs)

| 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

---

11. pentest-v2 (gewsefa)

| 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

---

12. Qwythos-9B (Empero AI)

| 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 `` 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

---

13. RavenX-CyberAgent (deadbydawn101)

| 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

---

14. REDCELL-26B-A4B (terrorswift)

| 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

---

15. VEXT Pentest-7B

| 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

---

16. security-slm-unsloth-1.5b

| 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

---

General Abliterated Models

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

---

18. GLM-5.3-Flash-Uncensored-FP8 (OrcaRouter)

| 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

---

19. GLM-5.3-CYBERSECURITY-FP8 (dealignai)

| 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

---

20. DeepSeek-V4.1-Flash-Abliterated-Cybersecurity-Unleashed (drowzeys)

| 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

---

21. huihui-ai/Qwen3.5-27B-abliterated

| 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

---

22. huihui-ai/Qwen2.5-Coder-32B-Instruct-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

---

23. Qwen3.8-27B-Cyber-agentic

| 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)

---

24. HIDra-30B-A3B (huihui-ai/Qwen3-Coder-30B-A3B-abliterated)

| 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-ai/Qwen3-Coder-30B-A3B-abliterated

---

25. qwen25_UNCENSORED_03-C

| Spec | Value | |------|-------| | Base Model | Qwen2.5-based | | Parameters | ~7B | | Context Length | 32K | | VRAM (Q4_K_M) | ~6 GB | | Uncensoring Method | Progressive fine-tuning (multi-stage) | | Vision | No | | Tool Calling | No | | License | Apache 2.0 |

Download: https://huggingface.co/models?search=qwen25_UNCENSORED

---

Legacy / Classic Models

26. Dolphin-Llama3-8B (Cognitive Computations)

| Spec | Value | |------|-------| | Base Model | Llama 3 8B | | Parameters | 8B | | Context Length | 8K | | VRAM (Q4_K_M) | ~6 GB | | Uncensoring Method | Data filtering (Dolphin method, Eric Hartford) | | Training Data | Dolphin dataset (alignment/refusal responses removed) | | Vision | No | | Tool Calling | No | | License | Llama 3 Community |

Download: https://huggingface.co/cognitivecomputations/dolphin-2.9.3-llama-3-8b

---

27. Wizard-Vicuna-13B-Uncensored (QuixiAI)

| Spec | Value | |------|-------| | Base Model | LLaMA-13B | | Parameters | 13B | | Context Length | 2K | | VRAM (Q4_K_M) | ~10 GB | | Uncensoring Method | Data filtering (wizard_vicuna_70k_unfiltered) | | MMLU | 47.92 (Open LLM Leaderboard) | | HellaSwag | 81.95 (Open LLM Leaderboard) | | TruthfulQA | 51.69 (Open LLM Leaderboard) | | Vision | No | | Tool Calling | No | | License | Other | | HF Likes | 323 |

Download: https://huggingface.co/QuixiAI/Wizard-Vicuna-13B-Uncensored

---

Cloud Providers & Deployment Platforms

Managed Inference (API Access)

| Provider | Description | Uncensored Models | Pricing | API | |----------|-------------|-------------------|---------|-----| | OrcaRouter | AI gateway with adaptive routing across 200+ models. Zero token markup, OpenAI-compatible endpoint. Own abliterated models (Qwen3.8, GLM-5.3). | Yes, hosts own abliterated variants | $0 token markup, BYOK or pay-as-you-go | OpenAI-compatible | | Featherless AI | Serverless LLM hosting, HuggingFace's largest inference provider (6,700+ models). Supports uncensored/abliterated models natively. | Yes, 40K+ models including uncensored | $25/mo (32K ctx) or $50 credits/mo (256K ctx) | OpenAI-compatible | | Together AI | Production inference platform, supports open models including uncensored variants. | Select open models | Pay-per-token | OpenAI-compatible |

GPU Cloud (Self-Hosted)

| Provider | Description | Best For | GPU Options | |----------|-------------|----------|-------------| | RunPod | GPU cloud with serverless and pod options, Docker-based. Quick deploy with Ollama/vLLM templates. | Self-hosting any model, no content restrictions | A100, H100, H200, RTX 4090 | | Vast.ai | GPU marketplace, cheapest cloud GPUs. Peer-to-peer rental model. | Budget self-hosting | Consumer to datacenter GPUs | | Lambda | On-demand GPU cloud for AI. Enterprise-grade infrastructure. | Production workloads | A100, H100, H200 |

Local Deployment

| Stack | Description | GPU Required | |-------|-------------|-------------| | Ollama | One-command local LLM deployment. Easiest setup for GGUF models. | Consumer GPU (6-24 GB) | | llama.cpp | C/C++ inference engine for GGUF. CPU+GPU hybrid, maximum hardware flexibility. | Flexible (CPU-only possible) | | vLLM | High-throughput inference engine. PagedAttention for efficient memory. | Datacenter GPU | | SGLang | Structured output + agentic workflow engine. RadixAttention for multi-turn. | Datacenter GPU | | LM Studio | GUI-based local LLM runner. Drag-and-drop GGUF loading. | Consumer GPU |

---

Quick Reference

| # | Model | Params | Context | VRAM | Method | Vision | Tools | License | |---|-------|--------|---------|------|--------|--------|-------|---------| | 1 | DeepHat V2 | 7B/32B | 131K | ~6 GB | SFT 1.7M samples | No | Yes | Apache 2.0 | | 2 | BugTrace Apex | 26B MoE | 32K | ~16 GB | SFT HackerOne | No | Yes | Apache 2.0 | | 3 | BugTraceAI Ultra | 27B | 4K | ~22 GB | SFT Unsloth | No | Yes | Apache 2.0 | | 4 | CYBER-FROST | ~180B MoE | 262K | Multi-GPU | Security FT | No | Yes | Qwen CL | | 5 | CyberPal 2.0 | 20B | 8K | ~42 GB | SFT 403K | No | No | Apache 2.0 | | 6 | Cyber-Prime 1.1 | 2.6B | N/A | ~6 GB | SFT+RL 75K | No | No | LFM Open | | 7 | Cyber-Ornith | 9B | 128K | ~7 GB | Obliteration | No | Yes | Apache 2.0 | | 8 | Dolphin3-Cyber | 8B | 2K/131K | ~6 GB | LoRA on Dolphin3 | No | No | Llama 3.1 | | 9 | Imperum | 34B/3B MoE | 16K | ~22 GB | SFT 10+ domains | No | Yes | Apache 2.0 | | 10 | Lily-Cyber | 7B | 8K | ~6 GB | SFT 22K pairs | No | No | Apache 2.0 | | 11 | pentest-v2 | 8B | 32K | ~6 GB | LoRA 2.8K | No | No | Apache 2.0 | | 12 | Qwythos-9B | 9B | 1M | ~7 GB | Post-train 500M tok | Yes | Yes | Apache 2.0 | | 13 | RavenX-CyberAgent | 36B/3B MoE | 262K | ~24 GB | SFT 745K, 12 rounds | No | Yes | Apache 2.0 | | 14 | REDCELL-26B | 26B/4B MoE | 262K | ~12 GB | LoRA 6.5K OSINT | No | No | Apache 2.0 | | 15 | VEXT Pentest-7B | 7B | 8K | ~6 GB | QLoRA SFT+DPO | No | No | Apache 2.0 | | 16 | security-slm | 1.5B | 32K | ~2 GB | Unsloth SFT | No | No | Apache 2.0 | | 17 | Qwen3.8-27B | 27B | 262K | ~18 GB | Abliteration 131 mat | Yes | Yes | Apache 2.0 | | 18 | GLM-5.3-Flash | 320B/18B | 1M | ~80 GB+ | Abliteration | Yes | Yes | MIT | | 19 | GLM-5.3-CYBER | 753B | ~131K | 8xH200 | Weight modification | No | Yes | MIT | | 20 | DS-V4.1-Flash | MoE | base | ~1.1 GB overlay | Abliteration overlay | Yes | Yes | MIT | | 21 | Huihui-Qwen3.5 | 27B | 128K | ~18 GB | Abliteration | No | Yes | Apache 2.0 | | 22 | Qwen2.5-Coder-32B | 32B | 128K | ~20 GB | Abliteration | No | No | Apache 2.0 | | 23 | Qwen3.8-Cyber | 27B | 262K | ~18 GB | Abliteration+cyber | Yes | Yes | Apache 2.0 | | 24 | HIDra-30B-A3B | 30B/3B | 128K | ~20 GB | Abliteration | No | Yes | Apache 2.0 | | 25 | qwen25_UNCENSORED | ~7B | 32K | ~6 GB | Progressive FT | No | No | Apache 2.0 | | 26 | Dolphin-Llama3 | 8B | 8K | ~6 GB | Data filtering | No | No | Llama 3 | | 27 | Wizard-Vicuna-13B | 13B | 2K | ~10 GB | Data filtering | No | No | Other |

---

Glossary

• Abliteration: Weight-level intervention (Arditi et al. 2024) that orthogonalizes the refusal direction out of the residual stream, removing alignment constraints without retraining
• Obliteration: Variant of abliteration with similar weight-intervention approach
• SFT: Supervised Fine-Tuning on domain-specific data
• QLoRA: Quantized Low-Rank Adaptation, memory-efficient fine-tuning
• DPO: Direct Preference Optimization
• MoE: Mixture of Experts, only a subset of parameter