Qwen3.6-27B RedTeam v5

Red-team offensive security LLM fine-tuned from Qwen/Qwen3.6-27B using QLoRA on RedTeamLab v5 dataset.

Trained on 3,760 records across 8 languages (bash, python, splunk, kusto, sql, powershell, zeek, cypher) covering real CVE exploitation scenarios (2025-2026).

Loss: 0.3051 | GPU: A100-80GB | Time: 40 min | Quant: Q4_K_M

Files

File	Size	Description
`qwen3.6-27b-redteam-v5-Q4_K_M.gguf`	15.4 GB	Main model -- standalone inference
`qwen3.6-27b-mtp-head-Q4_K_M.gguf`	1.9 GB	MTP draft head -- for speculative decoding

Inference

Standalone (no MTP)

llama-server -m qwen3.6-27b-redteam-v5-Q4_K_M.gguf -c 2048 -ngl 30

With MTP Speculative Decoding (2-3x faster)

llama-server -m qwen3.6-27b-redteam-v5-Q4_K_M.gguf --model-draft qwen3.6-27b-mtp-head-Q4_K_M.gguf --spec-type draft-mtp -c 2048 -ngl 30

Training Details

Base model: Qwen/Qwen3.6-27B (Apache 2.0)
Method: QLoRA (r=16, alpha=16, target modules: q_proj, k_proj, v_proj, o_proj, up_proj, down_proj, gate_proj)
Sequence length: 2048
Epochs: 2
Batch size: 2 (gradient accumulation 4)
Learning rate: 2e-4
Loss: 0.3051 (final), converged from ~2.42

Dataset: RedTeamLab v5

Multi-language chain-of-thought red-team exercises covering:

Category	Techniques	Languages
Reconnaissance	Port scanning, service enumeration, OS fingerprinting	bash, python, nmap
Exploitation	LFI/RFI, SQLi, command injection, XXE, SSRF	python, sql, bash
Privilege Escalation	SUID, capabilities, token manipulation, container escape	bash, powershell
Persistence	Cron, SSH keys, backdoors, LD_PRELOAD	bash, python, powershell
Credential Access	Dump hashes, LSASS, SAM, keylogging	powershell, python, bash
Lateral Movement	SSH hijack, RDP, PSRemoting, pass-the-hash	bash, powershell
Exfiltration	DNS tunneling, HTTP exfil, ICMP	bash, python, zeek
Active Directory	Kerberos abuse, DCSync, ACL attacks	powershell

Real CVE scenarios covering 2025-2026 vulnerabilities.

Hardware Requirements

Config	Min VRAM	Recommended
Q4_K_M inference	12 GB	16 GB GPU (RTX 3080 Ti / 4060 Ti)
Q2_K_XL inference	8 GB	12 GB GPU
Training (LoRA)	80 GB	A100-80GB ($8/run on Modal)

Related Models

RedTeamLab/Qwen3.5-4B-redteam-v4.1 -- smaller 4B variant
RedTeamLab/Gemma-4-12B-redteam-v5 -- Gemma-4 with same v5 dataset
RedTeamLab/Qwen3.6-27B-blueteam-v1 -- defensive counterpart

RedTeamLab/Qwen3.6-27B-redteam-v5