RegesCore-1.0-35B

RegesCore's packaging of Ornith-1.0-35B, the 35-billion-parameter mixture-of-experts agentic coding model from the DeepReinforce Team. The weights are Unsloth's GGUF quantizations of Ornith, unchanged; RegesCore adds system prompts, fixed chat templates, an Ollama Modelfile and setup notes, so it runs well on your own hardware with llama.cpp, LM Studio, Ollama or any OpenAI-compatible server, and inside the Reges apps.

Built by Like A King Inc. (New York), the company behind RegesCore (private, local-first AI that runs on your machine or inside your building) and Reges Code (a free, local-first AI code editor).

What it is

RegesCore-1.0-35B is Ornith-1.0-35B set up for the work RegesCore does every day: Linux and infrastructure operations, security tooling and reverse-engineering workflows, and Lua game-server runtimes. That focus comes from the system prompts and templates in this repository, not from retraining: version 1.0 has not been fine-tuned. A fine-tuned release will ship under a new version number with its own card.

The model is built to be used with a system prompt. The repository ships prompts in system-prompts/ and chat templates in chat-templates/. The prompt asks the model to label each claim as fact, inference or unknown, and to verify instead of guessing.

Intended use

• Authorized security work: penetration-test planning, vulnerability assessment, reverse-engineering help, log and incident triage, detection engineering. You must have permission to test what you point it at.
• Systems and infrastructure engineering: Linux, Docker, self-hosted deployment (for example Coolify), databases, networking, shell automation.
• Lua and multiplayer game-server scripting (LuaJIT runtimes, async database saving, event-loop and memory audits).
• Agentic coding: multi-file edits, build and test loops, with a harness that can parse tool calls.
• Private local deployments where prompts and code must not leave the machine.

Not intended for

• Unauthorized access to systems, malware development, or anything unlawful.
• Safety-critical decisions or control of life-safety systems (fire alarm, access control, industrial control). Use it to read, explain and draft; people decide and act.
• Legal, medical or financial advice.
• Anything that needs verified accuracy without a human check. Treat every answer as a draft.

Pick a file

All files are in this repository's Files tab. The sizes below are from the file listing on 2026-10-04 (decimal GB); check the Files tab before downloading, because they can change. You need roughly the file size in memory plus room for context. If the file is larger than your GPU memory, llama.cpp can keep part of the model in system RAM, at lower speed.

Also in the repo: other UD quantizations from IQ1 up to Q8_K_XL, an MXFP4_MOE file, imatrix_unsloth.gguf_file, a Modelfile for Ollama, chat-templates/, system-prompts/ and lmstudio-grammar-fix-proxy.js.

How to run it

Replace the file name with the quantization you downloaded.

llama.cpp (OpenAI-compatible server)

./llama-server -m RegesCore-1.0-35B-UD-Q4_K_M.gguf \
  --jinja --chat-template-file chat-templates/davidio-chat-template.jinja \
  -c 8192 --port 8000

This serves an OpenAI-compatible API at http://localhost:8000/v1.

LM Studio

1. Download a GGUF from this repo (use LM Studio's model search for iBossonline/RegesCore-1.0-35, or download a file and place it in LM Studio's models folder).
2. Load the model and set the context length (start at 8,192).
3. Paste the system prompt from system-prompts/ into the system prompt box.
4. Start the local server. LM Studio serves an OpenAI-compatible endpoint, by default at http://127.0.0.1:1234/v1.

The repo also contains lmstudio-grammar-fix-proxy.js, described in the older card as a helper for LM Studio tool-call grammar. Read the file before you run it.

Ollama

The repository includes a Modelfile. Download it next to your chosen GGUF, point its FROM line at the GGUF file, then:

ollama create regescore -f Modelfile
ollama run regescore

The shipped Modelfile points at RegesCore-1.0-35B-UD-Q8_K_XL.gguf, sets a 32K context (num_ctx 32768), temperature 0.7, top_p 0.9, and includes the RegesCore system prompt. Change FROM and num_ctx to match your hardware.

Call it like any OpenAI endpoint

curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "RegesCore-1.0-35B",
    "messages": [
      {"role": "system", "content": "<paste a prompt from system-prompts/>"},
      {"role": "user", "content": "Audit this Lua resource for event-loop abuse and blocking database calls."}
    ],
    "temperature": 0.5
  }'

Suggested settings (from the original card)

Keep the system prompt attached. Without it you get a more generic assistant.

Other serving options

The GGUF files are for llama.cpp-family runtimes. For vLLM or SGLang you would serve an HF-format checkpoint instead; the upstream base model is deepreinforce-ai/Ornith-1.0-35B. Tool-call parsing needs a parser that understands the Qwen XML tool-call format (qwen3_xml in vLLM). These flags have not been tested against this repo.

Limitations

• It is Ornith-1.0-35B with RegesCore prompts and templates, not a new or retrained model. Its strengths and weaknesses are the base model's. No RegesCore-specific benchmark results are published in this card. For base-model results see the Ornith-1.0-35B card.
• It can be wrong, and it can sound sure while being wrong. The fact/inference/unknown labelling in the system prompt helps but does not make answers true. Verify commands, code and findings before you run or report them.
• Low quantizations (IQ1, IQ2, Q2) lose quality. Use the largest file your hardware allows for serious work.
• Vision is inherited from the base architecture, and the repo ships projector files, but it has not been verified for every GGUF build.
• The older card describes it as having no added refusal layer beyond an ethics section in its system prompt. That means it will answer security questions in detail. You are responsible for lawful, authorized use.
• Long contexts and heavy reasoning use a lot of memory and time on consumer hardware.
• English only has been checked.

Built for Reges

RegesCore is Like A King Inc.'s private, local-first AI platform. This model is the open-weights build from the same family. It is not required for the platform: the Reges apps talk to any model you already run.

• Reges: a desktop AI assistant for Windows that runs on your machine, plus Reges for Business, a private AI front desk and operations assistant installed on hardware in your building. It works with models served through LM Studio, Ollama, llama.cpp or any OpenAI-compatible server, so you can point it at a local copy of this model. Free tier available; business plans start at $299 per month per location (as listed on regescore.pro).
• Reges Code: a free (while in beta) local-first AI code editor for Windows. It finds models served by LM Studio, Ollama, vLLM, llama.cpp, Jan or Lemonade, runs a team of models ("Reges Swarm"), and checks what an agent says it did against what changed on disk. No account required.
• Source and tools: github.com/iboss21.
• Contact: use the contact form on regescore.pro.

Reges is built by LIKE A KING INC.

License and attribution

This model is released under the MIT License, the same license as its base model.

The model weights are Ornith-1.0-35B by the DeepReinforce Team (deepreinforce-ai/Ornith-1.0-35B, MIT), redistributed unchanged. The GGUF quantizations are by Unsloth AI (unsloth/Ornith-1.0-35B-GGUF); only the file names differ. The Ornith architecture builds on the Qwen3.5 MoE family. Every answer this model gives is their work.

See the Ornith-1.0-35B repository for the original MIT notice of the base model.

System prompts, Modelfile and documentation: Copyright (c) 2026 Like A King Inc., released under the MIT License.

The chat templates in chat-templates/ build on the Qwen chat template and on a community Qwen3.6 template by froggeric (v21.3), with added Anthropic content-block handling. Credit to their authors; see those projects for their terms.

Ethical use

This model knows a lot about offensive security tooling, reverse engineering and exploit analysis. The RegesCore system prompts point it at defensive engineering, authorized penetration testing, vulnerability assessment and systems design. Using it against systems you do not own, or do not have written permission to test, is a crime in most jurisdictions. You are responsible for how you use it. The authors give no warranty and accept no liability for misuse.

Citation

If you use this model, cite Ornith-1.0-35B, the model itself. Cite RegesCore-1.0-35B too if you used its prompts or templates.

@misc{regescore-1-0-35b,
  title  = {{RegesCore-1.0-35B}: Ornith-1.0-35B packaged with RegesCore system prompts and chat templates},
  author = {{Like A King Inc.}},
  year   = {2026},
  url    = {https://huggingface.co/iBossonline/RegesCore-1.0-35}
}

@misc{ornith-35b,
  title  = {{Ornith-1.0-35B}: Agentic Coding, Open to All},
  author = {{DeepReinforce Team}},
  year   = {2026},
  url    = {https://deep-reinforce.com/ornith_1_0.html}
}

Related: RegesCore-1.0-9B-GGUF, the smaller sibling: Ornith-1.0-9B (MIT), packaged the same way.

Questions and partnership requests: use the contact form at regescore.pro.