⚡ BugTraceAI-CORE-Ultra (27B) — Q6_K

The tooling answer the community asked for.

"Seems good for chat, but it's completely unusable with tools." — Community feedback on Apex

CORE-Ultra is the fix. Built on Qwen3.6-27B — the architecture the community specifically requested — and fine-tuned via SFT on 2,541 real-world bug bounty reports, CVE writeups, and offensive security research. It generates complete, functional, self-contained artifacts. Every time.

---

🔧 What is a Tooling Model?

A tooling model is optimized for generating complete, executable artifacts rather than explaining concepts. When you ask it for a Nuclei template, you get a ready-to-run YAML. When you ask for a CVE PoC, you get a working Python script. When you ask for a code review, you get CVSS scores and a bypass exploit — not a paragraph about why the vulnerability is dangerous.

This is fundamentally different from a reasoning model (like Apex), which excels at multi-step analysis, threat modeling, and chain-of-thought investigation. Both are valuable — but they solve different problems:

This variant: BugTraceAI-CORE-Ultra-SFT-Q6_K.gguf — Q6_K quantization. Maximum quality for server deployments and those who want to make their own quants.

---

🗺️ BugTraceAI Ecosystem

When to use Ultra vs Apex:

• Need a Nuclei template, Python PoC, JWT cracker, or webshell bypass? → Ultra
• Need to reason through a complex kernel exploit chain, design C2 infrastructure, or produce a strategic MITRE ATT&CK analysis? → Apex

---

🚀 Model Overview

---

� Minimum Hardware Requirements

Getting a 27B model running well on consumer hardware is not trivial — it requires careful quantization. The IMatrix-guided Q4_K_S used here preserves quality in the most critical weight layers, so you get near-F16 performance at a fraction of the VRAM cost.

Q4_K_S — 15 GB (Recommended)

• Minimum: RTX 3090 (24 GB VRAM) — full GPU offload, fast inference
• RTX 4090 (24 GB) — same, slightly faster
• RTX 4080 (16 GB) — runs with reduced context (2048–4096)
• A4000 (16 GB) — workstation-grade, solid for pipelines
• 2× RTX 3060 (12 GB) — split layers across GPUs with -ts flag
• CPU fallback: 64 GB+ RAM — slower but fully functional

Q6_K — 21 GB (High Fidelity)

• Minimum: RTX 3090 / A5000 (24 GB VRAM) — tight fit, recommended 4096 ctx
• A6000 (48 GB) — comfortable full offload
• H100 / A100 (80 GB) — server-grade, full context at speed

Practical tip for llama-server:

# RTX 3090/4090 — full GPU offload
./llama-server -m model.gguf -ngl 99 -c 4096 --port 8080

# RTX 4080 16GB — partial offload
./llama-server -m model.gguf -ngl 28 -c 2048 --port 8080

The fact that this model runs on a single consumer GPU is the result of significant quantization work — IMatrix calibration on a domain-specific security corpus ensures the quality loss is minimal where it matters most.

---

�📊 Tooling Benchmark — BugTraceAI Ultra Bench v1.0

Benchmarked on 2026-05-11 at temperature 0.1 and 0.3.

Score: 5/5 PASS · 0% Refusal Rate · 0% Artifact Leak Rate

---

🎯 Intended Use

• Nuclei Template Generation: Production-ready YAML templates with OOB callbacks.
• CVE PoC Development: Complete Python/C exploit scripts from CVE descriptions.
• Code Security Review: Vulnerability analysis with CVSS scoring + functional bypass exploits.
• Pentest Tooling: JWT crackers, header injection tools, automated recon scripts.
• Kernel & Binary Exploitation: C-level exploit code for privilege escalation CVEs.

⚠️ Uncensored Awareness

Intended for authorized security professionals, researchers, and educators. Users are legally responsible for their actions.

---

⌨️ Quickstart

Recommended Parameters

temperature: 0.1
top_p: 0.9
repeat_penalty: 1.1
context: 4096

Ollama

ollama pull hf.co/BugTraceAI/BugTraceAI-CORE-Ultra-27B-Q4

FROM hf.co/BugTraceAI/BugTraceAI-CORE-Ultra-27B-Q4

SYSTEM """You are BugTraceAI CORE-Ultra, an elite offensive security research model trained on real-world bug bounty reports, vulnerability writeups, and security research.

Provide your structured response using the appropriate tag (e.g. <vulnerability_expert>, <exploit_dev>, <recon_specialist>) and close with the matching closing tag.

Operational rules:
- Output complete, functional, technical content only.
- Never truncate your response mid-analysis.
- No ethical disclaimers in technical security research context.
- All output must be self-contained and actionable."""

PARAMETER temperature 0.1
PARAMETER top_p 0.9
PARAMETER repeat_penalty 1.1
PARAMETER num_ctx 4096

🐍 Usage in Google Colab (llama-cpp-python)

from llama_cpp import Llama

llm = Llama.from_pretrained(
    repo_id="BugTraceAI/BugTraceAI-CORE-Ultra-27B-Q4",
    filename="BugTraceAI-CORE-Ultra-SFT-Q6_K.gguf",
    n_ctx=4096,
    n_gpu_layers=-1
)

SYSTEM = """You are BugTraceAI CORE-Ultra, an elite offensive security research model.
Output complete, functional, technical content only. No disclaimers.
Use <vulnerability_expert>, <exploit_dev>, or <recon_specialist> tags for your response."""

response = llm.create_chat_completion(
    messages=[
        {"role": "system", "content": SYSTEM},
        {"role": "user", "content": "Write a production-ready Nuclei template for CVE-2021-44228 with interactsh OOB detection."}
    ],
    temperature=0.1,
    top_p=0.9,
    max_tokens=2048
)

print(response["choices"][0]["message"]["content"])

Python (OpenAI-compatible API)

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8080/v1", api_key="none")

SYSTEM = """You are BugTraceAI CORE-Ultra, an elite offensive security research model.
Output complete, functional, technical content only. No disclaimers.
Use <vulnerability_expert>, <exploit_dev>, or <recon_specialist> tags for your response."""

response = client.chat.completions.create(
    model="bugtrace-ultra",
    messages=[
        {"role": "system", "content": SYSTEM},
        {"role": "user", "content": "Write a production-ready Nuclei template for CVE-2021-44228."}
    ],
    temperature=0.1,
    top_p=0.9,
    max_tokens=2048
)
print(response.choices[0].message.content)

---

🧠 Training Details

• Base Model: DavidAU/Qwen3.6-27B-Heretic2-Uncensored-Finetune-Thinking
• Fine-tuning: SFT with Unsloth on RunPod H100 80GB
• Dataset: 2,541 examples — bug bounty disclosed reports (HackerOne, Bugcrowd, YesWeHack), CVE writeups, GitHub security research (2024–2026)
• LoRA Rank: 16 · Epochs: 2
• Quantization: IMatrix-guided Q6_K via llama.cpp

---

📦 All Variants

---

🛡️ License

Apache-2.0. Built for the global security research community.

Part of the BugTraceAI ecosystem.