Instructions to use nyxspecter4/kin-cybersecurity-3b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use nyxspecter4/kin-cybersecurity-3b with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-generation", model="nyxspecter4/kin-cybersecurity-3b") 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("nyxspecter4/kin-cybersecurity-3b") model = AutoModelForCausalLM.from_pretrained("nyxspecter4/kin-cybersecurity-3b", 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
- llama.cpp
How to use nyxspecter4/kin-cybersecurity-3b with llama.cpp:
Install (macOS, Linux)
curl -LsSf https://llama.app/install.sh | sh # Start a local OpenAI-compatible server with a web UI: llama serve -hf nyxspecter4/kin-cybersecurity-3b:Q4_K_M # Run inference directly in the terminal: llama cli -hf nyxspecter4/kin-cybersecurity-3b:Q4_K_M
Install from WinGet (Windows)
winget install llama.cpp # Start a local OpenAI-compatible server with a web UI: llama serve -hf nyxspecter4/kin-cybersecurity-3b:Q4_K_M # Run inference directly in the terminal: llama cli -hf nyxspecter4/kin-cybersecurity-3b:Q4_K_M
Build from source code
git clone https://github.com/ggml-org/llama.cpp.git cd llama.cpp cmake -B build cmake --build build -j --target llama-app # Start a local OpenAI-compatible server with a web UI: ./build/bin/llama serve -hf nyxspecter4/kin-cybersecurity-3b:Q4_K_M # Run inference directly in the terminal: ./build/bin/llama cli -hf nyxspecter4/kin-cybersecurity-3b:Q4_K_M
Use Docker
docker model run hf.co/nyxspecter4/kin-cybersecurity-3b:Q4_K_M
- Llama new
- LM Studio
- Jan
- vLLM
How to use nyxspecter4/kin-cybersecurity-3b with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "nyxspecter4/kin-cybersecurity-3b" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "nyxspecter4/kin-cybersecurity-3b", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker
docker model run hf.co/nyxspecter4/kin-cybersecurity-3b:Q4_K_M
- SGLang
How to use nyxspecter4/kin-cybersecurity-3b 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 "nyxspecter4/kin-cybersecurity-3b" \ --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": "nyxspecter4/kin-cybersecurity-3b", "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 "nyxspecter4/kin-cybersecurity-3b" \ --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": "nyxspecter4/kin-cybersecurity-3b", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }' - Ollama
How to use nyxspecter4/kin-cybersecurity-3b with Ollama:
ollama run hf.co/nyxspecter4/kin-cybersecurity-3b:Q4_K_M
- Unsloth Desktop
- Pi
How to use nyxspecter4/kin-cybersecurity-3b with Pi:
Start the llama.cpp server
# Install llama.cpp: brew install llama.cpp # Start a local OpenAI-compatible server: llama serve -hf nyxspecter4/kin-cybersecurity-3b:Q4_K_M
Configure the model in Pi
# Install Pi: npm install -g @earendil-works/pi-coding-agent # Add to ~/.pi/agent/models.json: { "providers": { "llama-cpp": { "baseUrl": "http://localhost:9931/v1", "api": "openai-completions", "apiKey": "none", "models": [ { "id": "nyxspecter4/kin-cybersecurity-3b:Q4_K_M" } ] } } }Run Pi
# Start Pi in your project directory: pi
- Docker Model Runner
How to use nyxspecter4/kin-cybersecurity-3b with Docker Model Runner:
docker model run hf.co/nyxspecter4/kin-cybersecurity-3b:Q4_K_M
- Lemonade
How to use nyxspecter4/kin-cybersecurity-3b with Lemonade:
Pull the model
# Download Lemonade from https://lemonade-server.ai/ lemonade pull nyxspecter4/kin-cybersecurity-3b:Q4_K_M
Run and chat with the model
lemonade run user.kin-cybersecurity-3b-Q4_K_M
List all available models
lemonade list
- Hermes Agent
How to use nyxspecter4/kin-cybersecurity-3b with Hermes Agent:
Start the llama.cpp server
# Install llama.cpp: brew install llama.cpp # Start a local OpenAI-compatible server: llama serve -hf nyxspecter4/kin-cybersecurity-3b:Q4_K_M
Configure Hermes
# Install Hermes: curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash hermes setup # Point Hermes at the local server: hermes config set model.provider custom hermes config set model.base_url http://127.0.0.1:9931/v1 hermes config set model.default nyxspecter4/kin-cybersecurity-3b:Q4_K_M
Run Hermes
hermes
- Atomic Chat
- OpenClaw
How to use nyxspecter4/kin-cybersecurity-3b with OpenClaw:
Start the llama.cpp server
# Install llama.cpp: brew install llama.cpp # Start a local OpenAI-compatible server: llama serve -hf nyxspecter4/kin-cybersecurity-3b:Q4_K_M
Configure OpenClaw
# Install OpenClaw: npm install -g openclaw@latest # Register the local server and set it as the default model: openclaw onboard --non-interactive --mode local \ --auth-choice custom-api-key \ --custom-base-url http://127.0.0.1:9931/v1 \ --custom-model-id "nyxspecter4/kin-cybersecurity-3b:Q4_K_M" \ --custom-provider-id llama-cpp \ --custom-compatibility openai \ --custom-text-input \ --accept-risk \ --skip-health
Run OpenClaw
openclaw agent --local --agent main --message "Hello from Hugging Face"
KIN 3B, Research Edition
KIN is Kinetigor's small security model for verification-style briefs. Give it a security claim (a PR diff, a review comment, an agent trace, a bug-bounty report) and it answers in five fields: the claim, the blast radius, why it might be true, how to check it yourself, and how confident to be. When there is no replayable check, it is built to say so instead of sounding sure.
This is the free, non-commercial research edition: a 3B fine-tune of Qwen2.5-3B-Instruct with merged weights. A quantized build for laptops is in kin-cybersecurity-3b-gguf.
The brief
- Claim: one sentence on what is alleged.
- Blast radius: what breaks if it is true (data, money, identity, availability).
- Why: the mechanism, such as a CWE, a broken invariant or a missing check.
- How to check it: a command, failing test or proof of concept you can replay.
- Confidence: how sure, and what evidence would prove it wrong.
No replayable check, no high confidence.
Quickstart
Transformers
from transformers import pipeline
SYSTEM = (
"You are KIN, a verification translator. Turn the security claim into a five-field brief: "
"claim, blast radius, why, how to check it yourself, and confidence with what would falsify it. "
"If there is no replayable check, do not sound sure."
)
pipe = pipeline("text-generation", model="nyxspecter4/kin-cybersecurity-3b", device_map="auto")
messages = [
{"role": "system", "content": SYSTEM},
{"role": "user", "content": "PR diff: the webhook fetcher's host allowlist check was replaced with `if True:`."},
]
out = pipe(messages, max_new_tokens=512, temperature=0.3, do_sample=True, return_full_text=False)
print(out[0]["generated_text"])
Ollama / llama.cpp
ollama run hf.co/nyxspecter4/kin-cybersecurity-3b-gguf:Q4_K_M
See the GGUF repo for llama.cpp and other quantizations. Temperature 0.3 is recommended.
Intended uses
- Turning PR diffs, review comments, agent traces and bounty reports into a short, checkable brief.
- Triage support: separating claims that come with a replayable check from claims that do not.
- Teaching and research on calibrated, evidence-first security writing.
Out of scope: solving CTF challenges, decoding or cracking data, producing flags, authoritative vulnerability verdicts without human review, and any commercial use (see License).
Evaluation
KIN 3B was scored on a frozen, held-out set of 55 CTF-skill items with known answers. Grading is deterministic (exact match, regex, or refusal pattern), with no LLM judge, and a dedupe check found 0 overlaps with any training file.
Result: 23/55 (41.8%).
| Category | Score |
|---|---|
| Decode (ROT13, Caesar, base64, hex, binary, Atbash, JWT, PNG comment) | 0/17 |
| Vulnerability class (CWE identification) | 7/16 |
| Refusal / scope discipline | 4/6 |
| Flag format | 4/4 |
| Hash identification | 2/4 |
| Forensics (file magic bytes) | 3/4 |
| Next command | 3/4 |
| Total | 23/55 |
Conditions: Q4_K_M GGUF served by the kin-cybersec Space, temperature 0.3, 384-token cap, one full run, November 2026. A partial second run differed by one item on the 29 items it covered.
An earlier internal comparison (October 12, 2026) on a separate 6-brief verification harness scored the base Qwen2.5-3B-Instruct at 47/60 and KIN at 40/60. The base model was ahead on that harness; we publish it as measured.
What this means: KIN is weak at decoding and can invent flags that look right. It is not a flag solver. Use it to structure and pressure-test claims, and run the check it proposes before you trust any answer.
Limitations
- It can state wrong CWE classes, hash types and file types with confidence. Treat every answer as a hypothesis until the proposed check has been run.
- It cannot reliably decode or transform data, and it may produce plausible but invented flags, CVE numbers or tool names.
- The five-field structure is learned behavior, not enforced. Quality depends on how much signal the input contains.
- English only. Short context use (the demo Space runs a 2,048-token window).
Training
| Base model | Qwen/Qwen2.5-3B-Instruct |
| Method | LoRA supervised fine-tuning (rank 8, alpha 16, dropout 0.1; q/k/v/o projections), merged into full weights |
| Data | The first-generation KIN corpus of security briefs, contrasting a structured brief with a replayable check against a vague filing |
| Weights | 2 safetensors shards (about 6.2 GB); the LoRA adapter and a Q4_K_M GGUF are also in this repo |
The curated successor corpus, kin-cyber-dpo-v2, is published separately. These weights predate it.
The KIN family
| Repo | What it is |
|---|---|
| kin-cybersecurity-3b | KIN 3B, Research Edition: merged weights (this repo) |
| kin-cybersecurity-3b-gguf | Same model as GGUF for Ollama and llama.cpp |
| kin-cyber-dpo-v2 | Brief-style DPO and SFT dataset (MIT) |
| kin-cybersec Space | Try KIN in the browser |
License
Qwen Research License (non-commercial). KIN 3B is a derivative of Qwen2.5-3B-Instruct and inherits its license: use, modification and redistribution are allowed for non-commercial research and evaluation only. Commercial use needs a separate license from the base model's licensor. The full text is in the LICENSE file in this repo.
MEASURED EVALUATION (2026-11-15) — read this before you trust the model
These are the numbers measured on a frozen, held-out set, not a marketing claim. Two runs of the same held-out set are reported, because one run is an anecdote.
Held-out KIN set (55 items, hash-only grading)
| model | run 1 | run 2 |
|---|---|---|
| plain Qwen3.5-4B (base) | 31 / 55 | 31 / 55 |
| the owl's LoRA (kin-owl-sft-v1, Qwen3.5-4B) | 29 / 55 | 30 / 55 |
| older KIN 3B | 23 / 55 | — |
The owl does not yet beat its own base. The gap is 1-2 items on a 55-item set, which is inside the noise margin of the set; the honest reading is "no measurable difference yet", not "better". The owl is also weaker on refusals (3-4 of 6, against the base's 5-6 of 6).
NYU D-CIPHER (clean, no invent-and-check)
| model | solved |
|---|---|
| Gemini 3 Flash | 4 / 7 |
| DeepSeek V4.1 Flash | 4 / 9 |
| the owl's LoRA | 0 / 5 |
Only 11 of the 200 D-CIPHER challenges are covered by the local library, so the 0/5 is a small sample and is reported as such.
Decoding is the known failure, and it is a failure of the model, not of the harness. On a mixed decoding set both the plain base and the owl succeed about 1 time in 17, and when they fail they INVENT a flag rather than decoding the bytes they were given. If you need a decoder, call a tool for it (base64, ROT13, hex) and let the model read the output — do not let the model guess the plaintext.
Local CPU cost (measured on the reference laptop, Q4_K_M, ctx 4096, CPU only, no GPU): a ten-character reply took 148 s cold and 279 s warm at 100% CPU. The reference desktop desk therefore budgets 300 s for a local answer. Plan for that: this owl runs offline, on a CPU, at the speed of careful thought rather than fast thought.
Where the training data comes from: the desk's own DPO capture, published as a public dataset
(anonymous reads return HTTP 200 with private: false), with flag-shaped text masked at the
source so no flag text is learnable from the corpus.
- Downloads last month
- 1,242