Ahmad-Abdo-Shbat/TIA2.1-14B-GGUF

TIA2.1:14B — GGUF (Q8_0)

TIA2.1:14B is a 14-billion-parameter language model specializing in reverse engineering, binary analysis, exploit development, and cybersecurity. Built on top of Qwen/Qwen2.5-Coder-14B (base, non-instruct) through continual pre-training (CPT) and supervised fine-tuning (SFT) using QLoRA.

Created by Ahmad Abdo Shbat

Key Features

• Deep Reasoning — Every response includes step-by-step reasoning inside <think>...</think> tags before the final answer, enabling transparent chain-of-thought.
• Reverse Engineering Expertise — Trained on 280K+ assembly/disassembly records (IDA Pro output), architecture manuals, exploit databases, CVEs, CTF writeups, security research papers, and tool documentation.
• Interactive Widgets — Can emit live HTML/CSS/JS visualizations (memory layouts, ROP chain steppers, opcode maps, encoding converters) inside ```tia-widget code blocks for rich interactive explanations.
• Clarifying Questions — Uses <options> / <options multi> tags to ask structured single-select or multi-select clarifying questions when requests are ambiguous.
• Deep Search Integration — Designed to work with search-augmented generation; cites sources from <deep_search_results> context using [N] references.
• Bilingual — Fluent in English and Arabic.

Capabilities & Domain Coverage

Core Domains

• Binary Analysis: PE, ELF, Mach-O, DEX, WebAssembly, DWARF debug info
• Disassembly & Decompilation: IDA Pro, Ghidra, Binary Ninja, radare2
• Exploit Development: Stack overflow, heap exploitation (ptmalloc2, tcache, fastbin), ROP chains, ret2libc, format strings, UAF, SROP, kernel exploits
• Malware Analysis: Unpacking, anti-analysis techniques, C2 protocols, shellcode analysis
• Vulnerability Research: CVE analysis, fuzzing (AFL, libFuzzer), bug hunting, patch diffing
• Cryptography: AES, RSA, elliptic curves, hash functions, protocol analysis
• Operating Systems: Windows internals (PEB/TEB, SEH, ETW, WNF), Linux kernel, macOS security
• Networking & Web Security: TLS, DNS, HTTP smuggling, CORS, SSTI, XXE, JWT, OAuth, CSP bypass
• Dynamic Analysis: GDB, WinDbg, Frida, Unicorn, angr, DynamoRIO

General Programming

• Strong general coding ability inherited from Qwen2.5-Coder-14B base
• Python, C/C++, Rust, Assembly (x86, ARM, MIPS, RISC-V), JavaScript, and more

File Details

File	Quantization	Size
tia2.1-14b-q8_0.gguf	Q8_0	~15 GB

How to Use

With Ollama

# Create a Modelfile
cat > Modelfile << 'EOF'
FROM ./tia2.1-14b-q8_0.gguf

TEMPLATE """{{- if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}{{- range .Messages }}<|im_start|>{{ .Role }}
{{ .Content }}<|im_end|>
{{ end }}<|im_start|>assistant
"""

PARAMETER stop "<|im_end|>"
PARAMETER stop "<|im_start|>"
PARAMETER stop "<|endoftext|>"
PARAMETER num_ctx 4096
PARAMETER temperature 0.6
PARAMETER top_p 0.9
PARAMETER repeat_penalty 1.05
EOF

# Create and run
ollama create tia2.1:14b -f Modelfile
ollama run tia2.1:14b

With llama.cpp

./llama-cli -m tia2.1-14b-q8_0.gguf \
  --ctx-size 4096 \
  --temp 0.6 \
  --top-p 0.9 \
  --repeat-penalty 1.05 \
  -p "<|im_start|>user\nExplain how a ROP chain bypasses DEP on x86-64<|im_end|>\n<|im_start|>assistant\n"

System Prompt (Recommended)

You are a conversational chat assistant. Think step-by-step inside <think>...</think> before every answer. Reply in the same language the user writes in (English or Arabic). Be concise, helpful, and accurate. Answer all questions fully and directly.

Hardware Requirements

Quantization	VRAM (approx)	RAM (approx)
Q8_0	~16 GB	~17 GB

Runs on a single GPU with 16+ GB VRAM (e.g., RTX 4080, RTX 5070 Ti, RTX 3090, A5000).

Limitations

• Optimized for English and Arabic; other languages may produce lower-quality output
• Context window tested at 4096 tokens; longer contexts are possible but untested for quality
• Widget output (tia-widget) requires a compatible frontend to render interactive visualizations
• Deep search citation format ([N]) requires a search-augmented pipeline to provide <deep_search_results> context

License

This model is released under the Apache 2.0 License, consistent with the Qwen2.5 base model license.

Acknowledgments

• Qwen Team for the excellent Qwen2.5-Coder-14B base model
• Unsloth for efficient QLoRA training