A domain-specific LLM for cybersecurity — built for ASOC, DFIR and detection-engineering work, and built to run on one machine, inside your own perimeter.

Fine-tuned from Qwen/Qwen3.6-35B-A3B, a Mixture-of-Experts model with 35B total parameters but only ~3B active per token. You get large-model domain depth at small-model inference cost: roughly 50 tokens/second on a single NVIDIA DGX Spark, from a 20 GB file.

Released free to the security community by IMPERUM B.V. www.IMPERUM.io, built in partnership with Alican Kiraz. https://huggingface.co/AlicanKiraz0

Files

File	Size	Notes
`Imperum-CybersecurityLLM-v1.0.Q4_K_M.gguf`	20 GB	Recommended. ~22 GB RAM/VRAM. Fastest.
`Imperum-CybersecurityLLM-v1.0.Q8_0.gguf`	34 GB	Near-lossless. ~36 GB RAM/VRAM.

Everything is inside the .gguf — weights, tokenizer and chat template. No other files are required.

Quick start

llama.cpp (OpenAI-compatible server)

bash

llama-server -m Imperum-CybersecurityLLM-v1.0.Q4_K_M.gguf \
  --port 8099 --ctx-size 16384 --n-gpu-layers 99 \
  --jinja --temp 0.3 --reasoning off

Then point any OpenAI client at http://localhost:8099/v1.

Ollama

bash

ollama create imperum -f Modelfile
ollama run imperum

The bundled Modelfile sets num_ctx 8192 — a safe default for laptops, since KV cache costs memory on top of the weights. Raise it to 16384 if you have the headroom.

LM Studio

Drop the .gguf into your models folder and load it. Set temperature to 0.3.

Three things to know before you use it

1. It is a reasoning model. The chat template opens a <think> block by default. If you give it a small max_tokens, the whole budget is consumed by reasoning and content comes back empty — which looks like a broken deployment and isn't. Either pass --reasoning off (as above) or allow ≥500 tokens. To get thinking back for a single request, pass "chat_template_kwargs": {"enable_thinking": true}.

2. Use temperature ~0.3 for security work. The stock default is 1.0, and at 1.0 it will invent plausible-looking CVE numbers and rule syntax. 0.3 is set server-side in the example above; clients can still override per request.

3. It self-identifies as "Qwen". The fine-tune did not change the identity response. Cosmetic, not a misdeploy.

Capabilities

Trained across the full SOC workflow. The examples under each heading were produced by this model on a live deployment (Q8_0, temperature 0.3) — not claimed from the training config.

🔍 Alert Triage and Investigation

Turns a raw alert into an analyst-ready assessment: what most likely happened, how confident to be, and what to check before calling it a true positive. Given a Windows event sequence — 4624 logon from an unusual source IP → 4672 SeDebugPrivilege → 7045 service install → 4688 cmd.exe — it reconstructs the intrusion narrative, states a confidence level, and names the single most useful next log source to pull. Handles the classic parent/child cases (winword.exe spawning encoded PowerShell) with the right containment ordering.

🎯 Detection Engineering

Drafts Sigma and YARA content with the correct log source, the correct Sysmon event ID and correct ATT&CK tags attached — a working skeleton for an engineer to review and harden, rather than a blank editor. Also explains why a detection fires and where its false positives will come from.

🗺️ Threat Intelligence and ATT&CK Mapping

Maps intrusion chains to MITRE ATT&CK tactics and technique IDs unprompted. Verified correct on T1547.001 (Registry Run Keys), T1053.005 (Scheduled Task), T1204.002 (Malicious File) and T1059.001 (PowerShell). Give it a narrative — phishing ISO → LNK → rundll32 → scheduled task → LSASS dump → SMB lateral movement → staged exfiltration — and it returns the mapped kill chain.

🧯 Incident Response

First-hour checklists with the sequencing that actually matters: isolate the host without powering it off, capture volatile memory before disk, preserve event logs before analysis, and establish root cause before cleanup. Covers ransomware, business email compromise and confirmed-intrusion scenarios.

🦠 Malware Analysis

Explains persistence mechanisms and how to hunt them — autostart extensibility points, registry run keys, services, scheduled tasks, Winlogon and IFEO debuggers — along with the tooling (Autoruns, Procmon, Regshot, RegRipper) and the sandbox discipline to do it safely.

☁️ Cloud, Container and Identity Security

Reasons about modern attack paths end to end. From a pod spec with hostPID: true, privileged: true and a mounted docker socket, it walks container RCE → node root → kubelet → full cluster compromise, then names the admission policy that blocks it. Covers Kubernetes, IAM and cloud misconfiguration review.

🐍 Secure Code Review

Identifies vulnerabilities in source, explains the exploitation path, and returns a fixed version. Correctly flags unsafe pickle.loads() on user-controlled input as remote code execution and rewrites it safely.

📋 Vulnerability Management and GRC

Prioritisation reasoning, remediation guidance, control mapping and the documentation side of security operations — the reporting work that consumes analyst hours without needing analyst judgement.

Intended use: an assistant that accelerates human analysts — drafting, summarising, explaining and triaging. Not an autonomous decision-maker, and not a substitute for review of anything it produces. Detection logic it writes should be reviewed exactly as carefully as human-written logic.

Model details

Property	Value
Architecture	`qwen3_5_moe` (Qwen3.5-MoE)
Parameters	34.66B total, ~3B active per token
Layers	40
Hidden size	2048
Experts	256 routed, 8 active per token, plus a shared expert
Attention	Hybrid — 3× linear attention, then 1× full attention, repeating
Attention heads	16 query / 2 key-value, head dim 256
Vocabulary	248,320
Precision	bfloat16, merged weights (not an adapter)
Base	`Qwen/Qwen3.6-35B-A3B`

Training

Supervised fine-tune with LoRA r=16, α=32 across all linear layers, merged into the base weights afterwards. One epoch, learning rate 7e-5 with cosine schedule and 3% warmup, bfloat16, effective batch size 16, 7,750+ optimizer steps, maximum sequence length 5,120. The vision tower was frozen throughout.

Training data covers SOC/SIEM operations, detection engineering, DFIR, malware analysis, threat intelligence, vulnerability management, cloud/Kubernetes/IAM/ AppSec, OT/ICS, GRC and authorized offensive security.

Hardware it was tested on

One NVIDIA DGX Spark (GB10 Grace Blackwell, 128 GB unified memory, sm_121), llama.cpp with all layers offloaded to GPU. ~50 tokens/second single stream; ~25 tokens/second per stream with four concurrent requests.

Safety and responsible use

This model is trained for defensive security work: detection, investigation, response and hardening. It also covers authorized offensive security (penetration testing, red teaming) as practised under engagement.

Use it lawfully and only against systems you are authorized to test. As with any LLM, verify its output before acting on it — particularly detection logic, which should be reviewed exactly as carefully as anything written by hand.

Credits

Built in partnership with Alican Kiraz, whose cybersecurity training work made the domain depth in this model possible.
Base model: the Qwen team at Alibaba (Qwen/Qwen3.6-35B-A3B, Apache-2.0).
Released by IMPERUM B.V.
Website: www.IMPERUM.io

License

Apache 2.0, inherited from the base model.

Citation

bibtex

@misc{imperum-cybersecurity-llm-v1,
  title  = {Imperum-CybersecurityLLM-v1.0},
  author = {IMPERUM B.V. and Kiraz, Alican},
  year   = {2026},
  note   = {Fine-tuned from Qwen/Qwen3.6-35B-A3B},
  url    = {https://huggingface.co/IMPERUM/Imperum-CybersecurityLLM-v1.0-GGUF}
}