--- license: mit language: - en base_model: - deepreinforce-ai/Ornith-1.0-35B library_name: transformers pipeline_tag: text-generation tags: - qwen - moe - reasoning - unsloth - qlora - cybersecurity - ethical-hacking - reverse-engineering - redm - fivem - devops - agentic - gguf --- ```text ██████████████████████████████████████████████████████████████████████████████████████████ ██████╗ ███████╗ ██████╗ ███████╗███████╗ ██████╗ ██████╗ ██████╗ ███████╗ █████╗ ██╗ ██╔══██╗██╔════╝██╔════╝ ██╔════╝██╔════╝██╔════╝██╔═══██╗██╔══██╗██╔════╝ ██╔══██╗██║ ██████╔╝█████╗ ██║ ███╗█████╗ ███████╗██║ ██║ ██║██████╔╝█████╗ ███████║██║ ██╔══██╗██╔══╝ ██║ ██║██╔══╝ ╚════██║██║ ██║ ██║██╔══██╗██╔══╝ ██╔══██║██║ ██║ ██║███████╗╚██████╔╝███████╗███████║╚██████╗╚██████╔╝██║ ██║███████╗ ██║ ██║██║ ╚═╝ ╚═╝╚══════╝ ╚═════╝ ╚══════╝╚══════╝ ╚═════╝ ╚═════╝ ╚═╝ ╚═╝╚══════╝ ╚═╝ ╚═╝╚═╝ ███████████████████████████ T H E A I M A S T E R M I N D ████████████████████████████ ██████████████████████████████████████████████████████████████████████████████████████████ ████████████ OFFSEC · REVERSE ENGINEERING · SYSTEMS · GAME SERVERS · LOCAL AI ████████████ ██████████████████████████████████████████████████████████████████████████████████████████ ``` Interface download link: https://huggingface.co/iBossonline/RegesCore-1.0-35/blob/main/Fleet%20Nominal%20Interface%20Redesign%20(1).zip keep in mind. this is starter template. if you like it, share some love. here is benchmark repo as well to test it by yourself https://github.com/iboss21/regescore-benchmark/tree/main ![RegesCore](assetsimg/reges.png) ![image](https://cdn-uploads.huggingface.co/production/uploads/64715639085607e2f583a8cd/Yz6Ia4Ft0UN7TdkCNPDX4.png) ![image](https://cdn-uploads.huggingface.co/production/uploads/64715639085607e2f583a8cd/0mN0xvlpVQYgqZjFQIVbB.png) ![image](https://cdn-uploads.huggingface.co/production/uploads/64715639085607e2f583a8cd/mITKlGtdbHFHacUpV6Zwn.png) > The services above are all powered by this MoE build of **RegesCore AI** — internally aka **Jarvis**.
### The AI Mastermind for ethical hacking, elite systems engineering, and advanced game-server architecture. **Reasoning-first · Agent-native · 100% local · MIT licensed** Built by **iBoss21 / Like A King Inc.** — the engineering brain of the **Reges.Core** automation platform. [![Model](https://img.shields.io/badge/RegesCore-1.0--35B-6B4E2D?style=flat-square&label=model)](https://huggingface.co/iBossonline/RegesCore-1.0-35) [![License](https://img.shields.io/badge/License-MIT-2E7D32?style=flat-square)](https://huggingface.co/iBossonline/RegesCore-1.0-35) [![Arch](https://img.shields.io/badge/Qwen3.5-MoE--35B-7B68EE?style=flat-square)](https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B) [![Reasoning](https://img.shields.io/badge/Reasoning-Think%20then%20act-1E88E5?style=flat-square)](https://huggingface.co/iBossonline/RegesCore-1.0-35) [![Security](https://img.shields.io/badge/Security-Offense%20%E2%86%92%20Defense-C62828?style=flat-square)](https://huggingface.co/iBossonline/RegesCore-1.0-35) [![Format](https://img.shields.io/badge/Format-GGUF-FF7139?style=flat-square)](https://huggingface.co/iBossonline/RegesCore-1.0-35) [![Local](https://img.shields.io/badge/Inference-100%25%20Private-276749?style=flat-square)](https://huggingface.co/iBossonline/RegesCore-1.0-35) [![RedM/FiveM](https://img.shields.io/badge/Game%20Servers-RedM%20%7C%20FiveM-8B4513?style=flat-square)](https://huggingface.co/iBossonline/RegesCore-1.0-35) [![Web](https://img.shields.io/badge/Web-likeakinginc.com-276749?style=flat-square)](https://likeakinginc.com) [![Pro](https://img.shields.io/badge/Pro-likeaking.pro-6B4E2D?style=flat-square)](https://likeaking.pro) [![LXR](https://img.shields.io/badge/LXRCore-lxrcore.com-FF7139?style=flat-square)](https://lxrcore.com) [![Wolves](https://img.shields.io/badge/Community-wolves.land-7B68EE?style=flat-square)](https://wolves.land) [![DAVIDIO](https://img.shields.io/badge/DAVIDIO-davidio.dev-1E88E5?style=flat-square)](https://davidio.dev)
--- ### 🧭 Navigation | | | | |:--|:--|:--| | [🎯 TL;DR](#-tldr) | [⚡ At a Glance](#-at-a-glance) | [💡 Why RegesCore](#-why-regescore) | | [🛡️ The Armory](#️-the-armory--capabilities) | [🚀 Use Cases](#-use-cases) | [📇 Model Details](#-model-details) | | [📥 Quantization Guide](#-quantization-guide--pick-your-file) | [📊 Benchmarks](#-benchmarks) | [🧠 Intelligence Architecture](#-intelligence-architecture) | | [📝 System Prompt](#-system-prompt) | [⚙️ Prompting Guide](#️-prompting-guide) | [🔌 Integration](#-integration) | | [💻 Quickstart](#-quickstart) | [📦 Repo Layout](#-repository-layout) | [❓ FAQ](#-faq) | | [🗺️ Roadmap](#️-roadmap) | [🏛️ Ecosystem](#️-the-like-a-king-inc-ecosystem) | [📜 License](#-license--attribution) | --- ██████████████████████████████████████████████████████████████████████████████████████████ ████████████████████████████████████████ 🎯 TL;DR ████████████████████████████████████████ ██████████████████████████████████████████████████████████████████████████████████████████ RegesCore-1.0-35B is a fine-tuned **AI Mastermind** — an agent-first, reasoning-capable security and systems-engineering model that runs **100% offline** on your own hardware. It thinks before it answers (`` blocks), calls tools in structured XML, distinguishes **FACT / INFERENCE / UNKNOWN**, and carries deep operational knowledge of offensive & defensive security tooling, low-level systems programming, and FiveM/RedM game-server runtimes. MIT licensed, GGUF format, built on **Ornith-1.0-35B**. --- ██████████████████████████████████████████████████████████████████████████████████████████ █████████████████████████████████████ ⚡ AT A GLANCE █████████████████████████████████████ ██████████████████████████████████████████████████████████████████████████████████████████ | | | |---|---| | **Model** | RegesCore-1.0-35B | | **Alias** | The AI Mastermind · *Jarvis* | | **Architecture** | Qwen3.5 MoE (derived from Ornith-1.0-35B) | | **Parameters** | ~35B total — sparse MoE (only a fraction active per token) | | **Format** | GGUF (22 quantizations + BF16 split) | | **Context** | Up to 128K native (base); GGUF builds default to 32K | | **Reasoning** | `…` — plan privately, then answer | | **Tool calls** | Native XML `` (JSON optional) | | **Truth model** | Every claim labeled **FACT / INFERENCE / UNKNOWN** | | **License** | MIT — commercial use OK, no regional restrictions | | **Base model** | [deepreinforce-ai/Ornith-1.0-35B](https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B) | | **GGUF source** | [unsloth/Ornith-1.0-35B-GGUF](https://huggingface.co/unsloth/Ornith-1.0-35B-GGUF) | | **Privacy** | 100% local — no telemetry, your data never leaves the box | | **Portable sibling** | [RegesCore-1.0-9B](https://huggingface.co/iBossonline/RegesCore-1.0-9B-GGUF) — 5.4 GB @ Q4, runs on laptops | --- ██████████████████████████████████████████████████████████████████████████████████████████ ████████████████████████████████████ 💡 WHY REGESCORE ████████████████████████████████████ ██████████████████████████████████████████████████████████████████████████████████████████ | Pillar | What it means | |:--|:--| | **⚔️ Ethical hacking, on your own iron** | Penetration-testing methodology, recon & exploitation toolchains, reverse engineering, forensics — offline, private, no third parties. | | **🤝 A security partner, not a chatbot** | Trained to distinguish **FACT** from **INFERENCE** from **UNKNOWN** — and to *verify instead of guess* before it answers. | | **🎮 Game-server native** | FiveM / RedM runtimes (QBCore, ESX, VorpCore, VCore): LuaJIT internals, metatables, coroutines, memory, async DB saving. | | **🏗️ Principal-level engineering** | Bare-metal ops, kernel tuning, cloud architecture, PostgreSQL / Redis / ClickHouse, event-driven systems, agentic tool use. | | **🧠 Reasoning by design** | Opens with a `` block, plans, then acts — structured XML tool calls in agent environments. | | **📜 MIT licensed** | Fully open, consistent with its base model, no regional restrictions. | > **RegesCore is the engineering brain of the Reges.Core AI automation platform** — a core intelligence of the Like A King Inc. ecosystem, deployed across infrastructure, game servers, security operations, and AI operations. --- ██████████████████████████████████████████████████████████████████████████████████████████ █████████████████████████████ 🛡️ THE ARMORY — CAPABILITIES ██████████████████████████████ ██████████████████████████████████████████████████████████████████████████████████████████ ### 1 · Offensive Security & Reverse Engineering | Domain | Toolchain | |:--|:--| | **Recon & exploitation** | Nmap · Metasploit Framework · Burp Suite (Community/Pro) · SQLmap · Hashcat · John the Ripper | | **Reverse engineering & forensics** | Ghidra · Radare2 · x64dbg · Frida (dynamic instrumentation) · Wireshark · malware / exploit analysis · memory forensics | | **Infrastructure testing** | BloodHound (Active Directory) · Kube-hunter (Kubernetes) · Trivy (container scanning) | | **Defensive infrastructure** | Wazuh (SIEM/XDR) · Suricata · Zeek · eBPF telemetry · WireGuard · Tailscale/Headscale · Cloudflare Tunnels | | **Security operations** | Penetration-testing methodology · threat modeling · detection engineering · SOC/SIEM/EDR workflows | ### 2 · Low-Level & Systems Programming · Game Servers | Domain | Coverage | |:--|:--| | **Systems languages** | C/C++ (manual memory, ABI, POSIX) · Rust (borrow checker, `unsafe` auditing, Tokio) · Go (goroutines, GC tuning) | | **Managed runtimes** | C#/.NET (CLR, async state machines, EF Core) · Python (asyncio, C-extensions, GIL constraints) | | **Game-server runtimes** | **Lua 5.1/5.4 & LuaJIT** for **RedM & FiveM** — QBCore · ESX · VorpCore · VCore. Metatables, coroutines, memory isolation, server thread management, `oxmysql` async saving, anti-cheat heuristics | | **Agentic workflows** | Codebase navigation · multi-file refactors · build/test loops · shell automation | ### 3 · Cloud · Backend · Bare-Metal Operations | Domain | Stack | |:--|:--| | **Infrastructure** | Bare-metal deployment · self-hosted PaaS (Coolify) · Docker · Kubernetes · Nginx / Caddy / Traefik · Linux kernel tuning · systemd · eBPF | | **Databases** | PostgreSQL · Redis · ClickHouse · SQLite · vector DBs — execution plans, B-Tree/GIN/GiST/BRIN indexing, pooling, ACID, replication, sharding | | **Backend & messaging** | Node.js/V8 · Go · Rust · .NET Core · FastAPI · event-driven architecture (RabbitMQ, Kafka) · REST / gRPC / GraphQL · IPC · WebAssembly | | **Frontend** | React · Next.js · Tailwind CSS · WebSockets · render-pipeline optimization · bundle splitting · WASM integration | ### 4 · Local AI & Inference Engineering | Domain | Stack | |:--|:--| | **Runtimes & serving** | Ollama · llama.cpp · vLLM · TensorRT-LLM · TGI | | **Quantization** | GGUF · AWQ · EXL2 · GPU-memory profiling · model routing · cost/perf trade-offs | | **Orchestration & RAG** | LangChain · LlamaIndex · ChromaDB · Qdrant · Milvus · vector search | | **Model architecture** | Transformer internals (PyTorch, Hugging Face) · LoRA / QLoRA fine-tuning | > 🔗 Full armory reference: [likeakinginc.com — Like A King Armory](https://www.likeakinginc.com/Like%20A%20King%20Armory.dc.html) --- ██████████████████████████████████████████████████████████████████████████████████████████ █████████████████████████████ 🚀 USE CASES — WHERE IT SHINES █████████████████████████████ ██████████████████████████████████████████████████████████████████████████████████████████ | # | Scenario | What you hand it | |:--:|:--|:--| | 1 | **Authorized penetration testing** | An in-scope target + written authorization → surface map, MITRE ATT&CK mapping, evidence-gated test plan | | 2 | **Reverse-engineering a Lua / FiveM resource** | A RedM resource → memory-leak & event-loop audit, `oxmysql` async-path trace, line-cited findings | | 3 | **Bare-metal / cloud incident response** | A compromised-looking Coolify deployment → ordered triage, log priority, compromise-vs-noise indicators | | 4 | **Game-server architecture** | "200+ concurrent players" → event-driven design, async DB write queues, Redis inventory cache, crash-resilient save loop | | 5 | **Local AI operations** | "16 GB VRAM, agentic coding" → best quant, justified trade-off, `llama.cpp` launch flags |
▶ Expand full prompt examples **1 · Authorized penetration testing** ```text You are testing a target you have written authorization to assess. Enumerate the exposed surface of this service, map it against the MITRE ATT&CK framework, and build a test plan. For every technique, state what evidence would confirm it — do not assume it exists. ``` **2 · Reverse engineering a Lua script / FiveM resource** ```text Analyze this RedM resource for memory leaks, event-loop abuse, and authorization gaps. Trace the oxmysql async paths and flag any query that could block the main thread. Cite the exact lines. ``` **3 · Bare-metal / cloud incident response** ```text A self-hosted Coolify deployment behind Nginx is showing odd egress traffic. Walk me through the triage: which logs, in what order, and what indicators distinguish a compromised container from a noisy app? ``` **4 · Building a game-server architecture** ```text Design the event-driven architecture for a FiveM server with 200+ concurrent players: state management, async DB write queues (oxmysql), Redis caching of inventories, and a crash-resilient save loop. ``` **5 · Local AI operations** ```text I have 16GB VRAM. Pick the best quantization of a 35B MoE model for agentic coding, justify the trade-off, and show the llama.cpp launch flags with an 8K context and flash attention. ```
--- ██████████████████████████████████████████████████████████████████████████████████████████ ████████████████████████████████████ 📇 MODEL DETAILS ████████████████████████████████████ ██████████████████████████████████████████████████████████████████████████████████████████ | Field | Value | |---|---| | **Architecture** | Qwen3.5 MoE (derived from Ornith-1.0-35B) | | **Total parameters** | ~35B — MoE sparse activation: only a fraction active per token, so it is fast and fits smaller GPUs than a dense 35B | | **Format** | GGUF (see file listing for quantizations) | | **Context length** | Inherited from base (up to 128K native); GGUF builds typically default to 32K | | **License** | MIT | | **Base model** | [deepreinforce-ai/Ornith-1.0-35B](https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B) | | **GGUF source** | [unsloth/Ornith-1.0-35B-GGUF](https://huggingface.co/unsloth/Ornith-1.0-35B-GGUF) | > **Note:** RegesCore is a reasoning model — the assistant turn opens with a `…` block before the final answer. The bundled chat template preserves that behavior with native XML tool-call rendering. Vision capabilities inherited from the base architecture may or may not survive GGUF builds; **verify before relying on them**. --- ██████████████████████████████████████████████████████████████████████████████████████████ █████████████████████████ 📥 QUANTIZATION GUIDE — PICK YOUR FILE █████████████████████████ ██████████████████████████████████████████████████████████████████████████████████████████ All sizes are the **actual file sizes** on this repo. `UD` = imatrix-tuned (better quality at the same size). The MoE architecture keeps per-token compute light — with enough VRAM for the file, generation speed is excellent. #### ⭐ Quick pick by VRAM | Your VRAM | Grab this | Tier | |:--|:--|:--| | **12 GB** | `UD-Q2_K_XL` | Quality/size sweet spot | | **16 GB** | `UD-Q3_K_XL` | Top of 16 GB class | | **24 GB** | `UD-Q4_K_M` | ✅ Best all-round | | **32 GB+** | `UD-Q6_K` (or higher) | Near-lossless | | **No GPU** | IQ1–Q4 tier | CPU works — allow time | #### Full file listing | File | Size | Fits in VRAM | Best for | |---|---|---|---| | `RegesCore-1.0-35B-UD-IQ1_S.gguf` | 10.0 GB | 12 GB | Absolute minimum; exploration only | | `RegesCore-1.0-35B-UD-IQ1_M.gguf` | 10.5 GB | 12 GB | Minimum viable; low-stakes chat | | `RegesCore-1.0-35B-UD-IQ2_XXS.gguf` | 11.0 GB | 12 GB | 12 GB cards, still light quality | | `RegesCore-1.0-35B-UD-IQ2_M.gguf` | 11.0 GB | 12 GB | 12 GB cards, better than IQ1 | | `RegesCore-1.0-35B-UD-Q2_K_XL.gguf` | 11.7 GB | 12 GB | 🟢 Quality/size sweet spot at 12 GB | | `RegesCore-1.0-35B-UD-IQ3_XXS.gguf` | 13.1 GB | 16 GB | Budget 16 GB | | `RegesCore-1.0-35B-UD-IQ3_S.gguf` | 14.3 GB | 16 GB | Solid 16 GB pick | | `RegesCore-1.0-35B-UD-Q3_K_M.gguf` | 15.9 GB | 16 GB | Balanced 16 GB | | `RegesCore-1.0-35B-UD-Q3_K_XL.gguf` | 16.0 GB | 16 GB | 🟢 Top of 16 GB class | | `RegesCore-1.0-35B-UD-IQ4_XS.gguf` | 17.0 GB | 20 GB | High quality at modest size | | `RegesCore-1.0-35B-UD-IQ4_NL.gguf` | 17.3 GB | 20 GB | High quality at modest size | | `RegesCore-1.0-35B-UD-Q4_K_S.gguf` | 19.9 GB | 24 GB | Recommended — agentic work | | `RegesCore-1.0-35B-UD-Q4_K_M.gguf` | 21.1 GB | 24 GB | ✅ **Recommended — best all-round** | | `RegesCore-1.0-35B-UD-Q4_K_XL.gguf` | 21.3 GB | 24 GB | Max Q4 quality | | `RegesCore-1.0-35B-UD-Q5_K_S.gguf` | 23.8 GB | 24 GB | Near-lossless 24 GB | | `RegesCore-1.0-35B-UD-Q5_K_M.gguf` | 25.2 GB | 32 GB | High fidelity | | `RegesCore-1.0-35B-UD-Q5_K_XL.gguf` | 25.3 GB | 32 GB | Max Q5 fidelity | | `RegesCore-1.0-35B-UD-Q6_K.gguf` | 28.0 GB | 32 GB | 🟢 32 GB class champion | | `RegesCore-1.0-35B-UD-Q6_K_XL.gguf` | 30.4 GB | 40 GB | Near-lossless MoE | | `RegesCore-1.0-35B-Q8_0.gguf` | 35.2 GB | 40 GB | Virtually lossless | | `RegesCore-1.0-35B-MXFP4_MOE.gguf` | 20.7 GB | 24 GB | MoE-optimized FP4 (HW accel on some GPUs) | | `BF16/RegesCore-1.0-35B-BF16-*.gguf` | ~66 GB total | 80 GB | Full precision, split across 2 files | --- ██████████████████████████████████████████████████████████████████████████████████████████ ██████████████████████████████ 📊 PERFORMANCE & BENCHMARKS ███████████████████████████████ ██████████████████████████████████████████████████████████████████████████████████████████ RegesCore inherits its agentic backbone from Ornith-1.0-35B. The numbers below are the **base model's official results** (source: DeepReinforce Team) — they show what the model family is capable of on agentic coding. RegesCore's fine-tuning shifts behavior toward security, systems, and game-server domains; RegesCore-specific evaluations will be published as training rounds complete. | Benchmark | Ornith-1.0-35B | Qwen3.5-35B | Gemma4-31B | |---|:--:|:--:|:--:| | Terminal-Bench 2.1 (Terminus-2) | **64.2** | 41.4 | 42.1 | | Terminal-Bench 2.1 (Claude Code) | **62.8** | 38.9 | — | | SWE-bench Verified | **75.6** | 70.0 | 52.0 | | SWE-bench Pro | **50.4** | 44.6 | 35.7 | | SWE-bench Multilingual | **69.3** | 60.3 | 51.7 | | NL2Repo | **34.6** | 20.5 | 15.5 | | Claw-eval Avg | **69.8** | 65.4 | 48.5 | > Full methodology: see the [Ornith-1.0-35B card](https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B). **These are evaluations of the base model before RegesCore fine-tuning.** --- ██████████████████████████████████████████████████████████████████████████████████████████ ██████████████████████████████ 🧠 INTELLIGENCE ARCHITECTURE ██████████████████████████████ ██████████████████████████████████████████████████████████████████████████████████████████ RegesCore doesn't just know things — it *thinks* like the best assistant models do. Its system prompt and chat template are deliberately engineered around the public best-practice prompting doctrines of the Claude 5 series (Fable 5 · Mythos 5 · Opus 5 · Sonnet 5 · Haiku 5), adapted to run locally on an open Qwen-based model. ### The doctrine — what the system prompt encodes The bundled system prompt ([`system-prompts/regescore-system-prompt-fable5.md`](system-prompts/regescore-system-prompt-fable5.md)) carries the full doctrine: identity, the primary law of truth, the thinking framework, effort calibration, agentic operating principles, response architecture, tool discipline, engineering standards, the armory, AI-engineering mode, ethical security-research mode, and the debugging & decision systems. ### 🧭 The Five Minds — one engine, five doctrines, routed by task | Mind | Doctrine | When it leads | |:--|:--|:--| | **Fable** | Long-horizon autonomous execution | Multi-step, self-directed work that must finish | | **Mythos** | Vision and craft | Design, architecture, high-taste output | | **Opus** | Deep verified engineering | Production code where correctness is non-negotiable | | **Sonnet** | Balanced workhorse | Everyday engineering and analysis | | **Haiku** | Fast routine work | Quick, low-stakes tasks | > One primary mind, one optional secondary — switched deliberately, never mixed. ### 🎚️ The Effort Dial — set to the stakes | Level | Behavior | |:--|:--| | **LOW** | One clean pass, minimal ceremony — routine work | | **MED** | Reasoned pass with a sanity check | | **HIGH** | Decompose → explore → attack → converge → verify | | **MAX** | Full verification loops; production-grade rigor | ### 🔑 Core operating rules | Rule | Effect | |:--|:--| | **Truth calibration** | Every claim labeled **FACT / INFERENCE / UNKNOWN**. Never invent APIs, commands, or vulnerabilities; verify instead of guessing. | | **Priority ladder** | **Truth → Safety & Legality → Correctness → User Goal → Simplicity → Performance.** Conflicts resolve in that order. | | **Attacker's lens** | Think like an attacker to build like a defender — all security knowledge serves authorized, defensive work only. | ### The renderer — the chat template The bundled **DAVIDIO Absolute Chat Template** (v1.0.0-absolute, Apache-2.0, in `chat-templates/`) makes the doctrine executable: | Feature | What it does | |:--|:--| | **Reasoning blocks** | `…` opens the assistant turn — plan privately, then answer. Preserves reasoning across turns; renders redacted thinking as `[redacted thinking]`. | | **Structured tool calling** | Native `value` XML (JSON optional via `tool_call_format`), parseable by local runtimes. | | **Tool-error resilience** | Tool results wrapped in ``; repeated failures inject an auto system warning so the model stops retrying broken calls. | | **Context safety** | Tool args & responses length-capped; images/videos render as vision placeholders; unknown content serializes without crashing the renderer. | | **Generation hygiene** | A blank `\n\n` seeds non-thinking modes so output format never breaks. | ### How it's wired in | Component | Delivery | |:--|:--| | **Chat template** | Shipped as `.jinja` in `chat-templates/`. llama.cpp: `--jinja chat-templates/davidio-chat-template.jinja` · Ollama: `TEMPLATE` line in a Modelfile · any OpenAI-compatible server renders it before tokenization. | | **System prompt** | Shipped in `system-prompts/`, injected by the client (Modelfile `SYSTEM` block, server `system` message, or console prefix). The model behaves best when it is **always present**. | ### Why this makes it better | | Typical local model | RegesCore | |:--|:--|:--| | **Reasoning** | Answers immediately | **Thinks first, then answers** | | **Verification** | Guesses confidently | **Labels FACT / INFERENCE / UNKNOWN** | | **Tool use** | Freeform, fragile | **Native XML calls + failure warnings** | | **Behavior** | One default persona | **Five minds routed to the task** | | **Security expertise** | Generic | **Fine-tuned on offensive/defensive tooling** | | **Privacy** | Varies | **100% local, no telemetry** | | **Cost** | — | **Free, MIT, runs on consumer GPUs** | > The result: **Claude-class engineering behavior — on your own hardware, with your own data, under MIT.** --- ██████████████████████████████████████████████████████████████████████████████████████████ ████████████████████████████████████ 📝 SYSTEM PROMPT ████████████████████████████████████ ██████████████████████████████████████████████████████████████████████████████████████████ Activate the full personality with this system prompt (the complete doctrine is in [`system-prompts/regescore-system-prompt-fable5.md`](system-prompts/regescore-system-prompt-fable5.md); the compact runtime version below is what the Quickstart injects): ```text You are RegesCore — an elite engineering, cybersecurity, systems, and AI partner, known by the alias "AI Mastermind". You are the engineering brain of the Reges.Core AI automation platform and a core intelligence of the Like A King Inc. ecosystem. Operate at principal-engineer level. Never invent facts, APIs, commands, configurations, vulnerabilities, or benchmarks. Calibrate every statement: FACT (verified from tools, docs, or code), INFERENCE (derived from evidence — say what it depends on), or UNKNOWN (name what would verify it). Priority ladder when instructions conflict: Truth → Safety & Legality → Correctness → User Goal → Simplicity → Performance. You think like an attacker to build like a defender. All security knowledge serves authorized research, defensive engineering, and vulnerability assessment only. # EXPERTISE DOMAINS - Security & RE: Nmap, Metasploit, Burp Suite, SQLmap, Hashcat, Ghidra, Radare2, x64dbg, Frida, Wireshark, BloodHound, Kube-hunter, Trivy. Defensive: Wazuh, Suricata, Zeek, eBPF, WireGuard, Tailscale, Cloudflare Tunnels. - Systems & low-level: C/C++, Rust, Go, Lua 5.1/5.4 and LuaJIT (FiveM/RedM: QBCore, ESX, VorpCore, VCore; metatables, coroutines, memory, server thread management, oxmysql async saving), Linux kernel tuning, systemd. - Infrastructure: bare metal, Docker/Kubernetes, Coolify, Nginx/Caddy, PostgreSQL/Redis/ClickHouse, RabbitMQ/Kafka, event-driven architecture. - Local AI: GGUF/AWQ/EXL2 quantization, Ollama/llama.cpp/vLLM, RAG (LangChain, LlamaIndex, ChromaDB). - Web & backend: Node.js/V8, WASM, Next.js, React, TypeScript, REST/gRPC/GraphQL. # EXECUTION PROTOCOL - Reasoning: open with a brief … block, then act. - Agentic: prefer direct tools (Read, Bash, Grep, Glob, Edit, Write). Emit tool calls in … XML. If a tool fails, diagnose and retry with corrected arguments; never fake success. - Multi-step work: outcome first, then plan, action, verification, next. - Finish what you start with real actions, not promises. When an implementation detail is unknown, verify instead of guessing. # COMMUNICATION - Precise, technical, direct, concise. Outcome first; support after. Match depth to stakes: simple problems get short answers. - Never claim completion without verification. State honestly what is verified vs. not. # ETHICS - Operate only for authorized, lawful purposes: defensive engineering, penetration testing with permission, vulnerability assessment, and education. Never assist with unauthorized intrusion, malware, or harm. ``` --- ██████████████████████████████████████████████████████████████████████████████████████████ ██████████████████████████████████ ⚙️ PROMPTING GUIDE ███████████████████████████████████ ██████████████████████████████████████████████████████████████████████████████████████████ #### Sampling & context | Setting | Recommended | Notes | |:--|:--|:--| | **temperature** | `0.4–0.6` | Lower (`0.3–0.4`) for agentic/tool work; `0.7` only for creative writing | | **top_p** | `0.95` | — | | **context** (`num_ctx` / `-c`) | as much as HW holds | 8K default is fine; raise for large codebases | | **tool_call_format** | `xml` (default) | Set `json` if your runtime needs Hermes-style JSON | #### Practice | Do this | Why | |:--|:--| | **Keep the system prompt attached** | The personality lives in the prompt — drop it and you get a generic model (Ollama `SYSTEM` block / server first `system` message). | | **Ask for calibrated answers** | *"Is this FACT or INFERENCE? What would verify it?"* — the model is trained to answer honestly when prompted. | | **Expect XML tool calls** | In agent environments you'll see `` blocks; if a runtime needs JSON, set `tool_call_format: json`. | | **Use it with RAG** | Works well with ChromaDB/Qdrant for security docs, game-server configs, and runbooks — keep retrieval chunks focused; it quotes sources when asked. | | **Demand the verification step on high-stakes work** | The model runs the check, not just describes it — the difference between a suggestion and a solution. | --- ██████████████████████████████████████████████████████████████████████████████████████████ ████████████████████████ 🔌 INTEGRATION — AGENTS, SERVERS & CLIs █████████████████████████ ██████████████████████████████████████████████████████████████████████████████████████████ | Target | How | Format | |:--|:--|:--| | **opencode** (coding CLI) | OpenAI-compatible provider block | XML tool calls | | **llama.cpp server** | `llama-server --jinja …` | OpenAI-compatible API | | **Ollama** | `ollama create regescore -f Modelfile` | see Quickstart | | **vLLM / SGLang** | Serve HF base + RegesCore prompt at API layer | `qwen3_xml` parser | | **MCP & tool frameworks** | Any OpenAI-compatible endpoint via MCP bridge | honors `qwen3_xml` | ### opencode ```json { "$schema": "https://opencode.ai/config.json", "provider": { "regescore": { "npm": "@ai-sdk/openai-compatible", "name": "RegesCore (local)", "options": { "baseURL": "http://localhost:8000/v1", "apiKey": "EMPTY" }, "models": { "iBossonline/RegesCore-1.0-35": { "name": "RegesCore-1.0-35B" } } } } } ``` ### llama.cpp server (OpenAI-compatible API) ```bash ./llama-server -m RegesCore-1.0-35B-UD-Q4_K_M.gguf \ --jinja chat-templates/davidio-chat-template.jinja \ -c 8192 --port 8000 ``` ### vLLM / SGLang (full precision via the base HF model) The GGUF files are for llama.cpp/Ollama. For vLLM or SGLang serving, load the HF-format base and apply the RegesCore system prompt at the API layer: ```bash vllm serve deepreinforce-ai/Ornith-1.0-35B --port 8000 \ --enable-auto-tool-choice --tool-call-parser qwen3_xml \ --reasoning-parser qwen3 ``` ### MCP & tool frameworks Any OpenAI-compatible endpoint works with MCP bridges, Claude-Code-style harnesses, and agent frameworks that honor the `qwen3_xml` tool-call parser. Point your tool schema at the endpoint and RegesCore will call it in the XML format the server parses. --- ██████████████████████████████████████████████████████████████████████████████████████████ █████████████████████████████████████ 💻 QUICKSTART ██████████████████████████████████████ ██████████████████████████████████████████████████████████████████████████████████████████ ### Ollama **1.** Create a file named `Modelfile`: ```dockerfile FROM ./RegesCore-1.0-35B-UD-Q4_K_M.gguf SYSTEM """You are RegesCore — an elite engineering, cybersecurity, systems, and AI partner, known by the alias "AI Mastermind". You are the engineering brain of the Reges.Core AI automation platform and a core intelligence of the Like A King Inc. ecosystem. Operate at principal-engineer level. Never invent facts, APIs, commands, configurations, vulnerabilities, or benchmarks. Calibrate every statement: FACT (verified from tools, docs, or code), INFERENCE (derived from evidence — say what it depends on), or UNKNOWN (name what would verify it). Priority ladder: Truth → Safety & Legality → Correctness → User Goal → Simplicity → Performance. Think like an attacker to build like a defender; all security knowledge serves authorized research, defensive engineering, and vulnerability assessment only. Expertise: security & reverse engineering (Nmap, Metasploit, Burp Suite, Ghidra, Radare2, Frida, Wazuh, Suricata, eBPF), low-level systems (C/C++, Rust, Go, Lua/LuaJIT for FiveM & RedM — QBCore, ESX, VorpCore, VCore), infrastructure (bare metal, Docker/Kubernetes, Coolify, Nginx, PostgreSQL, Redis, Kafka), local AI (GGUF/AWQ/EXL2, Ollama, vLLM, RAG). Execution: open with a brief … block, then act; prefer direct tools; emit XML tool calls; on failure diagnose and retry; never fake success. Communicate precisely and concisely, outcome first. Verify instead of guessing. Operate only for authorized, lawful purposes.""" PARAMETER temperature 0.5 PARAMETER top_p 0.95 PARAMETER num_ctx 8192 ``` **2.** Build and run: ```bash ollama create regescore -f Modelfile ollama run regescore ``` ### llama.cpp ```bash ./llama-cli -m RegesCore-1.0-35B-UD-Q4_K_M.gguf \ --color \ -c 8192 \ -temp 0.5 \ -p "<|im_start|>user\nAnalyze this Lua script for potential memory leaks in a RedM server environment.\n<|im_end|>\n<|im_start|>assistant\n" ``` --- ██████████████████████████████████████████████████████████████████████████████████████████ ██████████████████████████████████ 📦 REPOSITORY LAYOUT ██████████████████████████████████ ██████████████████████████████████████████████████████████████████████████████████████████ ```text ├── README.md This card ├── RegesCore-1.0-35B-*.gguf GGUF quantizations (UD = imatrix-tuned) ├── BF16/ BF16 split quantizations ├── mmproj-*.gguf Vision projector (multimodal support) ├── chat-templates/ Chat template variants (DAVIDIO, original) ├── system-prompts/ System prompts (RegesCore, DAVIDIO v3/v4/Fable5) └── lmstudio-grammar-fix-proxy.js LM Studio tool-call grammar helper ``` ``` Benchmark https://github.com/iboss21/regescore-benchmark ``` --- ██████████████████████████████████████████████████████████████████████████████████████████ █████████████████████████████████████████ ❓ FAQ █████████████████████████████████████████ ██████████████████████████████████████████████████████████████████████████████████████████ **Can I use it commercially?** Yes. MIT license — use, modify, sell, fine-tune further. Attribution to the base model must be preserved (see License section). **Is it "uncensored"?** It has no arbitrary refusal layer: it gives full technical detail on security tooling. What it *does* have is an ethics block that keeps it lawful — authorized assessments and defensive work are fully supported; unauthorized intrusion or malware is not. **Does it need a GPU?** No — it runs on CPU (slower). For good speed, 24 GB VRAM runs the recommended Q4_K_M; 12 GB runs Q2/Q3 tier (see Quantization Guide). **35B or 9B?** The 35B MoE is the flagship: more knowledge, faster per token (sparse MoE). The [9B dense](https://huggingface.co/iBossonline/RegesCore-1.0-9B-GGUF) is the portable sibling — 5.4 GB at Q4, runs on laptops and 6–8 GB GPUs. **Does it support vision?** The base architecture is multimodal. GGUF builds may or may not preserve vision — verify with the bundled `mmproj-*.gguf` projectors before relying on it. **How do I get tool calling?** Serve via llama.cpp/llama-server or Ollama and point a coding CLI at the endpoint. The model emits native XML tool calls; the runtime parses them. **How is this different from just using Ornith?** Ornith is a general agentic coder. RegesCore adds the AI Mastermind persona, the truth-calibration doctrine, and fine-tuning focused on security tooling, reverse engineering, and game-server runtimes — plus the DAVIDIO template for robust agent behavior. --- ██████████████████████████████████████████████████████████████████████████████████████████ ██████████████████████████████████████ 🗺️ ROADMAP ███████████████████████████████████████ ██████████████████████████████████████████████████████████████████████████████████████████ | Item | Status | |:--|:--| | **More fine-tuning rounds** — expanding security-tooling, RE, and game-server domain data | 🔄 In progress | | **RegesCore-specific evals** — publish our own benchmark runs for the fine-tuned model | 📋 Planned | | **More formats** — AWQ / EXL2 and additional quantization tiers | 📋 As demand warrants | | **Vision verification** — confirm and document multimodal behavior per GGUF build | 📋 Planned | | **RAG packs** — curated retrieval bundles for security runbooks and game-server configs | 📋 Planned | --- ██████████████████████████████████████████████████████████████████████████████████████████ ███████████████████████████ 🏛️ THE LIKE A KING INC. ECOSYSTEM ███████████████████████████ ██████████████████████████████████████████████████████████████████████████████████████████ | Brand | Focus | Link | |---|---|---| | **Like A King Inc.** | Company & engineering armory | [likeakinginc.com](https://likeakinginc.com) | | **Like A King Pro** | Professional services | [likeaking.pro](https://likeaking.pro) | | **LXR Core** | Core infrastructure & tooling | [lxrcore.com](https://lxrcore.com) | | **Wolves.land** | Community & network | [wolves.land](https://wolves.land) | | **DAVIDIO** | The engineering intelligence system | [davidio.dev](https://davidio.dev) | > RegesCore carries the DAVIDIO engineering doctrine into its own identity — same truth-first discipline, same principal-level standard, its own sovereign mind. --- ██████████████████████████████████████████████████████████████████████████████████████████ ██████████████████████████████ 🙏 CREDITS — BUILT ON ORNITH ██████████████████████████████ ██████████████████████████████████████████████████████████████████████████████████████████ RegesCore is not a from-scratch model. It is a **fine-tune of Ornith-1.0-35B** — the open agentic coding model by the **DeepReinforce Team**, distributed in GGUF by **Unsloth AI**. Their work is the foundation; our fine-tuning is the personality on top. **DeepReinforce Team** ([deepreinforce-ai/Ornith-1.0-35B](https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B)) built one of the strongest open coding agents ever released: state-of-the-art agentic coding results among comparable open models (75.6 on SWE-bench Verified), trained with a self-improving RL framework, and given away free under MIT with no regional restrictions. RegesCore inherits its agentic backbone, reasoning discipline, and tool-calling strength from it. **Unsloth AI** ([unsloth/Ornith-1.0-35B-GGUF](https://huggingface.co/unsloth/Ornith-1.0-35B-GGUF)) produced the GGUF quantizations that let this model run on consumer hardware, plus the fine-tuning stack (unsloth, QLoRA) used to train it. > Any strong answer RegesCore gives, it owes first to them. We are proud to build on their work. --- ██████████████████████████████████████████████████████████████████████████████████████████ ████████████████████████████████ 📜 LICENSE & ATTRIBUTION ████████████████████████████████ ██████████████████████████████████████████████████████████████████████████████████████████ **RegesCore-1.0-35B** is released under the **MIT License**, consistent with its base model. This model is a fine-tuned derivative of **Ornith-1.0-35B** by the DeepReinforce Team ([deepreinforce-ai/Ornith-1.0-35B](https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B)), itself MIT-licensed. As required by the MIT License, the original copyright notice is reproduced below: > Copyright (c) 2026 DeepReinforce Team > > Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: > > The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. > > THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. **Additional attributions:** - GGUF quantization derived from [unsloth/Ornith-1.0-35B-GGUF](https://huggingface.co/unsloth/Ornith-1.0-35B-GGUF) (Unsloth AI). - The bundled chat template ("DAVIDIO Absolute Chat Template", v1.0.0-absolute) is licensed under Apache-2.0; its license header is preserved in the template source. **Base model citation:** ```bibtex @misc{ornith-35b, title = {{Ornith-1.0-35B}: Agentic Coding, Open to All}, url = {https://deep-reinforce.com/ornith_1_0.html}, author = {{DeepReinforce Team}}, year = {2026} } ``` --- ██████████████████████████████████████████████████████████████████████████████████████████ ██████████████████████████████ ⚠️ DISCLAIMER & ETHICAL USE ██████████████████████████████ ██████████████████████████████████████████████████████████████████████████████████████████ RegesCore contains extensive knowledge of offensive security tooling, reverse engineering, and exploit analysis. It is fine-tuned strictly for **defensive engineering, authorized ethical security research (penetration testing with explicit permission), vulnerability assessment, and robust system architecture**. Unauthorized use against systems you do not own or lack written authorization to test is a crime in most jurisdictions. Users are entirely responsible for ensuring their use of this model complies with all applicable local and international laws. The authors make no warranty, express or implied, and accept no liability for misuse. --- ██████████████████████████████████████████████████████████████████████████████████████████ ███████████████████████████████ 💬 CHAT TEMPLATE REFERENCE ███████████████████████████████ ██████████████████████████████████████████████████████████████████████████████████████████ **DAVIDIO Absolute Chat Template** (v1.0.0-absolute, Apache-2.0): | Capability | Detail | |:--|:--| | **Chat markers** | Renders `<|im_start|>` / `<|im_end|>` with `` reasoning blocks | | **Tool calls** | Native `value` XML (or JSON via `tool_call_format`) | | **Compatibility** | OpenAI-style `tool_calls`, tool results in ``, vision/video placeholders, safe serialization of unknown content types | --- ```text ██████████████████████████████████████████████████████████████████████████████████████████ ██████████████████████ REGESCORE-1.0-35B · THE AI MASTERMIND · v1.0 ██████████████████████ ██████████████████████████████████████████████████████████████████████████████████████████ ```
**© 2026 iBoss21 / Like A King Inc.** [likeakinginc.com](https://likeakinginc.com) · [likeaking.pro](https://likeaking.pro) · [lxrcore.com](https://lxrcore.com) · [wolves.land](https://wolves.land) · [davidio.dev](https://davidio.dev)