RegesCore
The AI Mastermind — a reasoning-capable security, systems, and game-server engineering model, fine-tuned by Davidio and built by Like A King Inc. Think before it answers, verify instead of guess, and run entirely on your own iron.
RegesCore-1.0-35B
Flagship- ArchitectureQwen3.5 MoE (from Ornith-1.0-35B)
- Parameters~35B — sparse activation
- ContextUp to 128K native (GGUF defaults 32K)
- Sweet spot24 GB → UD-Q4_K_M (21.1 GB)
Quantization picks
- Best all-round: UD-Q4_K_M · 21.1 GB
- 12 GB cards: UD-Q2_K_XL · 11.7 GB
- 16 GB cards: UD-Q3_K_XL · 16.0 GB
- 32 GB+: UD-Q6_K · 28.0 GB
RegesCore-1.0-9B
Portable- ArchitectureQwen3.5 Dense (from Ornith-1.0-9B)
- Parameters~9B — dense
- ContextGGUF defaults to 32K
- Sweet spot6–8 GB → Q4_K_M (5.4 GB)
Quantization picks
- Recommended: Q4_K_M · 5.4 GB
- 8 GB strict: Q5_K_M · 6.2 GB
- Near-lossless: Q8_0 · 9.1 GB
- Full precision: bf16 · 17.1 GB
What RegesCore is trained for
Offensive Security & RE
- Nmap · Metasploit · Burp Suite
- SQLmap · Hashcat · John the Ripper
- Ghidra · Radare2 · x64dbg · Frida
- Wireshark · malware & memory forensics
Defensive Infrastructure
- Wazuh SIEM/XDR · Suricata · Zeek
- eBPF telemetry · threat modeling
- BloodHound AD auditing
- Kube-hunter · Trivy · WireGuard · Tailscale
Low-Level & Game Servers
- C/C++ · Rust · Go — memory, ABI, unsafe
- Lua 5.1/5.4 & LuaJIT for FiveM/RedM
- QBCore · ESX · VorpCore · VCore
- oxmysql async paths · server threads
Cloud, Backend & Local AI
- Bare metal · Coolify · Nginx · Caddy
- PostgreSQL · Redis · ClickHouse · Kafka
- GGUF/AWQ/EXL2 · Ollama · llama.cpp · vLLM
- RAG: LangChain · LlamaIndex · ChromaDB
Why it behaves like an engineer, not a chatbot
Truth calibration
Every claim is labeled FACT · INFERENCE · UNKNOWN. RegesCore verifies instead of guessing — never invents APIs, commands, or vulnerabilities.
Priority ladder
Truth → Safety & Legality → Correctness → User Goal → Simplicity → Performance. When instructions conflict, the model resolves them in that order.
Reasoning by design
Opens with a private <think> block, plans, then acts. Structured XML tool calls in agent environments — no proprietary APIs required.
100% local & private
Runs offline on your own hardware. No telemetry, no third parties, no data leaving your machine. MIT licensed — free to use, modify, and sell.
Inherited from the Ornith backbone
Official base-model results before RegesCore fine-tuning. RegesCore-specific evaluations are published as training rounds complete.
| Benchmark | 35B MoE | 9B Dense |
|---|---|---|
| SWE-bench Verified | 75.6 | 69.4 |
| Terminal-Bench 2.1 (Terminus-2) | 64.2 | 43.1 |
| SWE-bench Pro | 50.4 | 42.9 |
| SWE-bench Multilingual | 69.3 | 52.0 |
| Claw-eval Avg | 69.8 | 63.1 |
Ollama quickstart
# Modelfile FROM ./RegesCore-1.0-9B-Q4_K_M.gguf SYSTEM """You are RegesCore — the AI Mastermind...""" PARAMETER temperature 0.5 PARAMETER top_p 0.95 PARAMETER num_ctx 8192 ollama create regescore-9b -f Modelfile ollama run regescore-9b
llama.cpp server
./llama-server -m RegesCore-1.0-35B-UD-Q4_K_M.gguf \ --jinja chat-templates/davidio-chat-template.jinja \ -c 8192 --port 8000 # OpenAI-compatible, agent-ready: # native <tool_call> XML, <think> reasoning # blocks, qwen3_xml parser support
Built inside the Like A King Inc. ecosystem
Run the Mastermind on your own machine
Free, open, MIT-licensed. Download from Hugging Face and run fully offline — from a laptop with 6 GB of VRAM to a full 80 GB rig.