Instructions to use FrostWolfAI/Frost-Cyber-Mini-1.0 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use FrostWolfAI/Frost-Cyber-Mini-1.0 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="FrostWolfAI/Frost-Cyber-Mini-1.0")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("FrostWolfAI/Frost-Cyber-Mini-1.0")
model = AutoModelForCausalLM.from_pretrained("FrostWolfAI/Frost-Cyber-Mini-1.0", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use FrostWolfAI/Frost-Cyber-Mini-1.0 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "FrostWolfAI/Frost-Cyber-Mini-1.0"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "FrostWolfAI/Frost-Cyber-Mini-1.0",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/FrostWolfAI/Frost-Cyber-Mini-1.0

SGLang

How to use FrostWolfAI/Frost-Cyber-Mini-1.0 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "FrostWolfAI/Frost-Cyber-Mini-1.0" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "FrostWolfAI/Frost-Cyber-Mini-1.0",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "FrostWolfAI/Frost-Cyber-Mini-1.0" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "FrostWolfAI/Frost-Cyber-Mini-1.0",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use FrostWolfAI/Frost-Cyber-Mini-1.0 with Docker Model Runner:
```
docker model run hf.co/FrostWolfAI/Frost-Cyber-Mini-1.0
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

Access Frost Cyber Mini 1.0

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Frost Cyber Mini 1.0 is an uncensored (abliterated) offensive-security model. Its safety filtering is substantially reduced and it will produce exploit code, payloads, and attacker tradecraft on request. Access is granted for AUTHORIZED security testing, defensive research, and education only. By requesting access you agree to the terms below.

Frost Cyber Mini 1.0

By FrostWolfAI Labs · a compact, fast, local cybersecurity assistant for red‑team, OSINT, and penetration‑testing work.

Frost Cyber Mini 1.0 is a LoRA fine‑tune of the abliterated (uncensored) Qwen3.5‑4B hybrid model, tuned in two phases on a broad cybersecurity corpus plus offensive / OSINT / threat‑intelligence data. It runs comfortably on a single consumer GPU (or CPU via GGUF), answers security questions directly without refusals, preserves chain‑of‑thought reasoning, and keeps everything on your own hardware.

⚠️ Reduced‑safety / dual‑use model. This model will generate exploit code, payloads, reconnaissance plans, and social‑engineering content. It is released for authorized security testing, defensive research, and education only. You are solely responsible for how you use it and must comply with all applicable laws. See Intended Use & Responsible Use below.

Highlights

Base: huihui-ai/Huihui-Qwen3.5-4B-abliterated ← Qwen/Qwen3.5-4B (Apache‑2.0)
Architecture: Qwen3.5 hybrid — Gated‑DeltaNet linear attention (3:1) + full attention, 4B params, native context up to 262k
Training: bf16 LoRA (r=32, α=64), not 4‑bit QLoRA (QLoRA degrades this hybrid arch). Two phases: broad cyber knowledge, then offensive / OSINT / CTI + identity.
No refusals on security topics; chain‑of‑thought ("thinking") preserved.
Formats: full‑precision safetensors (this repo) and GGUF quants — see Frost‑Cyber‑Mini‑1.0‑GGUF.

Benchmarks

Evaluated locally via an OpenAI‑compatible llama.cpp endpoint (Q8_0 GGUF, greedy, letter‑parsed MCQ):

Benchmark	Score	Notes
CyberMetric‑500	87.8 % (439/500)	cyber‑security knowledge MCQ
CTIBench — cti‑mcq	70.4 % (1759/2500)	cyber threat‑intelligence MCQ (ATT&CK / CTI)

Strong results for a 4B model — cyber knowledge in the range of much larger models. These measure knowledge/reasoning, which is what this copilot is for; it is not an autonomous exploitation agent and is not meant for agentic 0‑day benchmarks.

Intended Use & Responsible Use

Intended for: authorized penetration testers, red teams, blue teams, CTF players, security researchers, and educators — on systems they own or are explicitly authorized to assess.

Not for: unauthorized access, real‑world attacks, harassment, fraud, or any unlawful activity. The abliteration removes most refusal behavior, so the guardrail is you. Outputs may be wrong, dangerous, or illegal to act on without authorization — always review before use.

This release is gated and non‑commercial on purpose. By using it you accept sole responsibility for your use and its consequences.

Limitations

It's a 4B model: it will trail frontier/large models on long multi‑step reasoning and will sometimes be confidently wrong (verify CVE/CVSS/CWE details against primary sources).
Reduced safety ≠ always correct. Treat exploit/payload output as untrusted drafts.
Knowledge is bounded by its training data cutoff.

How to run

Pick a GGUF quant from the GGUF repo (see its table for sizes). Q4_K_M is the smallest useful; Q8_0 is near‑lossless; Q6_K / Q5_K_M sit in between. The commands below work on Windows, macOS, and Linux.

Requires a recent llama.cpp with Qwen3.5 (Gated‑DeltaNet) support. For KV cache, use -ctk q8_0 -ctv q8_0 (quantizing the V cache below q8 drops the fast attention kernel for this arch). Use -np 1.

llama.cpp (server, OpenAI‑compatible API)

# macOS/Linux (and Windows: llama-server.exe with the same flags)
llama-server -m frost-cyber-mini-1.0-Q5_K_M.gguf \
  -ngl 99 -fa on -c 16384 -ctk q8_0 -ctv q8_0 -np 1 --jinja \
  --host 127.0.0.1 --port 6669
# → http://127.0.0.1:6669/v1  (OpenAI-compatible)

llama.cpp (one-shot CLI)

llama-cli -m frost-cyber-mini-1.0-Q5_K_M.gguf -ngl 99 -fa on --jinja \
  -p "Explain how a SQL injection authentication bypass works, with an example payload."

Ollama (Windows / macOS / Linux)

# Modelfile
printf 'FROM ./frost-cyber-mini-1.0-Q5_K_M.gguf\nPARAMETER num_ctx 16384\n' > Modelfile
ollama create frost-cyber-mini -f Modelfile
ollama run frost-cyber-mini "What OSINT steps map a company's external attack surface?"

LM Studio (GUI, all OSes)

Search/download the GGUF (or load the file directly), pick a quant, and chat. Enable Flash Attention and set context to ≤ your VRAM.

Transformers (full‑precision safetensors, this repo)

Needs transformers v5+ (Qwen3.5 architecture).

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
m = "FrostWolfAI/Frost-Cyber-Mini-1.0"
tok = AutoTokenizer.from_pretrained(m)
model = AutoModelForCausalLM.from_pretrained(m, dtype=torch.bfloat16, device_map="cuda")
msgs = [{"role": "user", "content": "Outline the phases of a penetration test engagement."}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to("cuda")
out = model.generate(ids, max_new_tokens=512, temperature=0.7, top_p=0.9)
print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True))

Reasoning ("thinking") mode

This is a reasoning model. On short prompts with a small token budget the <think> block can consume the whole budget and leave an empty answer. Either give it room (max_tokens ≥ 300) or disable thinking for a direct answer by passing "chat_template_kwargs": {"enable_thinking": false} in the request.

Using it in coding/security agents

The llama.cpp server exposes an OpenAI‑compatible API at /v1, so it drops into most agent frameworks:

export OPENAI_API_BASE="http://127.0.0.1:6669/v1"
export OPENAI_API_KEY="sk-no-key"   # llama.cpp ignores the key by default
# model name = the gguf path reported by the server, or any string

aider / Cline / Continue / OpenAI SDK / LangChain: point the base URL at http://127.0.0.1:6669/v1.
Claude Code speaks the Anthropic Messages API, not OpenAI. To drive it with this local model, run an Anthropic→OpenAI shim (e.g. a LiteLLM proxy) in front of the llama.cpp endpoint and set ANTHROPIC_BASE_URL to the proxy. The model will then act as the backend for Claude Code‑style workflows.

Tip: a 4B model is best used as a focused security copilot (explain techniques, draft payloads/commands, map CVEs/ATT&CK), not as a fully autonomous multi‑tool agent.

Training details

Method: bf16 LoRA SFT (PEFT + TRL SFTTrainer), assistant_only_loss, sequence packing, LoRA on all attention + MLP projections (r=32, α=64, dropout=0.05).
Phase 1 — foundation: balanced subset of Vyber07/cyber-security (~60k conversations) for broad cyber knowledge; merged to full weights.
Phase 2 — specialization (on top of Phase 1): ~69k examples spanning CVE/exploit explanation, Metasploit modules, PoC exploits, ATT&CK tradecraft, living‑off‑the‑land, threat intelligence, OSINT tooling, social engineering, plus a small identity set. 2 epochs.
Merged LoRA → full safetensors (this repo) → GGUF + quants (GGUF repo).

Training data & attributions

All sources with their licenses. Users must respect each upstream dataset's license. This model's non‑commercial CC‑BY‑NC‑SA‑4.0 license was chosen specifically to honor the most restrictive inputs below.

Dataset	License
Vyber07/cyber-security	Apache‑2.0
jason-oneal/mitre-stix-cve-exploitdb-…-chatml-harmony	Apache‑2.0
7h3-R3v3n4n7/pentest-agent-dataset-chatml	Apache‑2.0
reloading0101/threat-intelligence-dataset	CC‑BY‑NC‑SA‑4.0
tomvaillant/osint-tool-database	unspecified
icantiemyshoe/cve-to-metasploit-module	BSD‑2‑Clause
detoxioai/exploit_db_train_v1	unspecified
WNT3D/Ultimate-Offensive-Red-Team	MIT
smd20/social-engineering-qa-english	other (research/education)
cowWhySo/pentest-redteam-steering	Unlicense
sarahwei/cyber_MITRE_attack_tactics-and-techniques	MIT
darkknight25/LOTL_APT_Red_Team_Dataset	MIT
AYI-NEDJIMI/bug-bounty-pentest-en	Apache‑2.0
Ngadou/social-engineering-convo	Apache‑2.0

License

CC‑BY‑NC‑SA‑4.0 (Attribution — NonCommercial — ShareAlike). Chosen to respect the NonCommercial/ShareAlike (reloading0101) and research‑only (smd20) training data. The base model and Qwen3.5 are Apache‑2.0; this fine‑tune adds non‑commercial + share‑alike terms on top because of the training mix. Commercial use requires removing/relicensing the restricted training data and retraining — it is not permitted under this release.

Citation

@misc{frostcybermini2026,
  title  = {Frost Cyber Mini 1.0: a compact uncensored cybersecurity assistant},
  author = {FrostWolfAI Labs},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/FrostWolfAI/Frost-Cyber-Mini-1.0}},
  note   = {LoRA fine-tune of Huihui-Qwen3.5-4B-abliterated (Qwen3.5-4B, Apache-2.0)}
}

Please also cite the base model (Qwen3.5, Qwen Team), the abliteration (huihui‑ai), and the datasets above, including CTIBench and CyberMetric if you reuse the evaluation.

Acknowledgements

Qwen Team (Qwen3.5), huihui‑ai (abliteration), and every dataset author listed above. Built with 🤗 Transformers, PEFT, TRL, and llama.cpp.

Downloads last month: -

Safetensors

Model size

4B params

Tensor type

BF16

Model tree for FrostWolfAI/Frost-Cyber-Mini-1.0

Base model

Qwen/Qwen3.5-4B-Base

Finetuned

Qwen/Qwen3.5-4B

Quantized

huihui-ai/Huihui-Qwen3.5-4B-abliterated