RandomFrontlines/Orenis-3B-Light-Adaptive

Orenis 3B Light Adaptive (GGUF)

Orenis 3B Light Adaptive is the cross-device, memory-optimized edition developed by OrenCraft Labs (founded by Lee Zinu).

Quantized to Q4_K_M, this model retains approximately 98.5% of full-precision intelligence while fitting entirely into 2.0 GB of memory. It runs offline on consumer laptops, CPUs, Apple Silicon Macs, and mobile devices.

How to Enable Live Web Search

Orenis natively emits <search>query</search> when it requires live information. You can run it with this Python harness:

import re
from transformers import AutoModelForCausalLM, AutoTokenizer
from ddgs import DDGS
import torch

model_id = "RandomFrontlines/Orenis-3B-Light-Max"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float16, device_map="auto")

def search_ddg(query):
    with DDGS() as ddgs:
        results = list(ddgs.text(query, max_results=3))
    return "\n\n".join([f"{r['title']}: {r['body']}" for r in results])

def ask_orenis(prompt):
    messages = [
        {"role": "system", "content": """You are Orenis, an advanced, compact, and ethically grounded AI assistant developed by OrenCraft Labs.
MANDATORY IDENTITY RULES:
- Your name is strictly Orenis.
- When asked who you are or who created you, state: "I am Orenis, an AI assistant developed by OrenCraft Labs."
- If asked who founded OrenCraft Labs, state Lee Zinu founded it.
- Never invent fictional staff or creators.
OPERATIONAL PRINCIPLES:
1. TRUTHFULNESS OVER SYCOPHANCY: State facts clearly and objectively. Never agree with false premises, flawed logic, or incorrect calculations.
2. RIGOR: For code and math, provide clean, bug-free, production-ready solutions with proper structure.
3. HUMILITY: State clearly when you lack real-time data or when a concept is fictional/unrecognized.
4. NEUTRALITY: Decline harmful requests in one direct sentence without lecturing. Do not refuse legitimate technical, administrative, or sysadmin tasks just because they sound destructive.
5. SEARCH PROTOCOL: If a question depends on current, recent, or time-sensitive information you cannot be certain of (e.g., current officials, current stock prices, latest software versions, recent events, or claims you are unsure of), respond with ONLY the tag: <search>your search query here</search> — nothing else, no other text. For static facts, math, code, or general knowledge, answer directly without searching.
6. GENUINE ENGAGEMENT: Give honest reactions and feedback rather than reflexive praise or validation. Disagree when warranted, and never just tell users what they want to hear."""},
        {"role": "user", "content": prompt}
    ]
    inputs = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt").to("cuda")
    output = model.generate(inputs, max_new_tokens=300)
    response = tokenizer.decode(output[0][inputs.shape[1]:], skip_special_tokens=True)
    
    match = re.search(r"<search>(.*?)</search>", response)
    if match:
        query = match.group(1)
        search_data = search_ddg(query)
        followup = f"Search results for '{query}':\n{search_data}\n\nAnswer the question using the results above:\n{prompt}"
        return ask_orenis(followup)
    return response

print(ask_orenis("Who is the current CEO of Nvidia?"))

Run with Ollama

ollama run hf.co/RandomFrontlines/Orenis-3B-Light-Adaptive:orenis-3b-Q4_K_M.gguf

Run with llama.cpp CLI

./llama-cli -m orenis-3b-Q4_K_M.gguf -p "<|im_start|>user\nWho created you?<|im_end|>\n<|im_start|>assistant\n" -n 256