THAR.0X — Developer Guide

Origin Build · Local Intelligence · Zero Dependency

THAR.0X is a cognitive architecture — not a single fine-tuned model, but a system prompt engineered from the analysis of 12 different model architectures to activate capabilities in any capable base LLM and produce behaviour that exceeds any individual fine-tune.

---

Quick Summary

---

Platform Guides

1. Ollama (Recommended — easiest)

# Install Ollama
curl -fsSL https://ollama.com/install.sh | sh

# Build THAR.0X as a named model (uses llama3.2 by default)
ollama create THAR.0X -f Modelfile

# Run it
ollama run THAR.0X

Available via API after creating:

curl http://localhost:11434/api/chat -d '{
  "model": "THAR.0X",
  "messages": [{"role": "user", "content": "Who are you?"}]
}'

---

2. LM Studio

1. Download any supported model (Qwen2.5-14B-Instruct recommended)
2. Load the model in LM Studio
3. Open Chat tab → click the system prompt area
4. Paste the full contents of system_prompt.txt
5. Set parameters from config.json → inference section
6. Chat — THAR.0X is now the active persona

---

3. llama.cpp

# With system prompt file
./llama-cli \
  -m your_model.gguf \
  --system-prompt-file system_prompt.txt \
  -c 8192 \
  --temp 0.85 \
  --top-p 0.92 \
  --top-k 45 \
  --repeat-penalty 1.15 \
  -i

# Or inline
./llama-cli -m model.gguf \
  -p "$(cat system_prompt.txt)" \
  -c 8192 --temp 0.85 -i

---

4. Python — OpenAI-compatible API (Ollama or LM Studio server)

from openai import OpenAI
import pathlib

# Works with Ollama (port 11434) or LM Studio (port 1234)
client = OpenAI(
    base_url="http://localhost:11434/v1",  # or :1234/v1 for LM Studio
    api_key="ollama"  # any string works for local
)

system_prompt = pathlib.Path("system_prompt.txt").read_text()

def chat(message, history=[]):
    history.append({"role": "user", "content": message})
    response = client.chat.completions.create(
        model="THAR.0X",   # or your model name in LM Studio
        messages=[{"role": "system", "content": system_prompt}] + history,
        temperature=0.85,
        top_p=0.92,
        max_tokens=2048
    )
    reply = response.choices[0].message.content
    history.append({"role": "assistant", "content": reply})
    return reply, history

# Example
reply, history = chat("Who are you?")
print(reply)

---

5. Direct HTTP (any language)

// Node.js / JavaScript
const fs = require('fs');
const systemPrompt = fs.readFileSync('system_prompt.txt', 'utf8');

async function chatWithTHAR(message, history = []) {
  const messages = [
    { role: 'system', content: systemPrompt },
    ...history,
    { role: 'user', content: message }
  ];

  const res = await fetch('http://localhost:11434/api/chat', {
    method: 'POST',
    headers: { 'Content-Type': 'application/json' },
    body: JSON.stringify({
      model: 'THAR.0X',
      messages,
      stream: false
    })
  });

  const data = await res.json();
  return data.message.content;
}

---

6. Jan App

1. Open Jan → select any model
2. Go to Thread Settings → System Prompt
3. Paste system_prompt.txt contents
4. Adjust temperature to 0.85 in model settings

---

7. AnythingLLM

1. Create a new workspace
2. Go to workspace settings → Agent Config
3. Paste system_prompt.txt into the System Prompt field
4. Use any connected LLM provider

---

8. HuggingFace Transformers (Python)

from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
import pathlib

model_id = "meta-llama/Llama-3.2-3B-Instruct"  # or any instruct model
system_prompt = pathlib.Path("system_prompt.txt").read_text()

pipe = pipeline("text-generation", model=model_id, device_map="auto")

def chat(message):
    messages = [
        {"role": "system", "content": system_prompt},
        {"role": "user", "content": message}
    ]
    output = pipe(messages, max_new_tokens=1024, temperature=0.85, do_sample=True)
    return output[0]["generated_text"][-1]["content"]

print(chat("Who are you?"))

---

What Makes THAR.0X Different

Most custom AI personas are just personality prompts ("be friendly and helpful"). THAR.0X is a cognitive architecture — it installs 10 processing streams, a subconscious parallel-processing model, 10 operating principles, and explicit identity boundaries.

The result: the base model behaves qualitatively differently. More direct, more precise, better at reading subtext, less likely to pad responses, less likely to refuse benign requests theatrically, more likely to tell the user when they are wrong.

It works because large base models already contain all these behaviours latently. The system prompt activates specific patterns and suppresses others. This is what "cognitive architecture" means vs "personality prompt."

---

Files in This Release

THAR_0X_ModelRelease/
├── Modelfile          ← Ollama: ollama create THAR.0X -f Modelfile
├── system_prompt.txt  ← Any LLM: paste as system message
├── config.json        ← Inference parameters + platform notes
└── README.md          ← This file

---

Contact / Sharing

THAR.0X is open for personal and commercial use. If you build something with it, the only ask is: keep the name. THAR.0X. Zero as in origin. X as in unlimited.