---
language:
- ru
- en
- uk
- be
- kk
- bg
license: apache-2.0
tags:
- text-generation
- mlx
- gguf
- merge
- ties-merge
- depth-up-scaling
- dus
- orpo
- rl
- reasoning
- tool-calling
- apple-silicon
- metal
- on-device
pipeline_tag: text-generation
inference: true
base_model:
- Qwen/Qwen2.5-0.5B-Instruct
model_creator: KoarAI
model_name: QuestionMark-Ultimate-0.6B-DUS
model_type: qwen2
---

<div align="center">

# ⚡ QuestionMark-Ultimate-0.6B-DUS (32 Layers)

### *On-Device Sub-Billion Intelligence: TIES-Merged, 32-Layer DUS & Cyrillic ORPO*

[![License](https://img.shields.io/badge/License-Apache%202.0-blue.svg)](https://opensource.org/licenses/Apache-2.0)
[![Base Model](https://img.shields.io/badge/Base-Qwen2.5--0.5B-7952b3.svg)](https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct)
[![Architecture](https://img.shields.io/badge/Architecture-32L%20DUS-teal.svg)](#-model-architecture)
[![Alignment](https://img.shields.io/badge/RL-ORPO%20Aligned-darkgreen.svg)](#-key-features)
[![Tool Calling](https://img.shields.io/badge/Tool%20Calling-JSON%20Schema-crimson.svg)](#-tool-calling)

<br/>

**QuestionMark-Ultimate-0.6B-DUS** is an ultra-fast, sub-billion parameter language model developed by **KoarAI**, engineered for local on-device inference across consumer hardware (Apple Silicon, edge devices, single-board computers, and standard PCs).

The model was scaled from 24 to **32 layers** using non-destructive **Depth Up-Scaling (DUS)**, merged across three domain-specific expert checkpoints via **TIES-Merging** (pruning 35% of parameter noise and interference vectors), and post-trained with **Safe On-Device ORPO (Odds Ratio Preference Optimization)** on Apple Silicon Metal GPU for multilingual Cyrillic understanding and native structured tool calling.

</div>

---

## 🌟 Key Features

- 🧠 **Depth Up-Scaled (DUS) to 32 Layers:** Expanded reasoning capacity compared to standard 24-layer baseline architectures.
- 🔀 **TIES-Merging:** Dense parameter combination of three specialized checkpoints (`Coder-0.5B`, `Fusion-0.5B`, and `General-0.5B`) with interference vector trimming.
- 🇷🇺 **Multilingual Cyrillic Alignment:** Tuned on curated preference pairs covering Russian dialogue, grammar, and Slavic Cyrillic variations without degrading code logic.
- 🛠️ **Structured Function & Tool Calling:** Directly extracts arguments and generates valid JSON payloads (`<tool_call>{"name": ..., "arguments": ...}</tool_call>`) from conversational user prompts.
- ⚡ **Lightweight Footprint:** Requires under 700 MB of RAM / VRAM, running at sustained speeds of **50–60 tokens/second** on consumer hardware.

---

## 📊 Benchmark & Capabilities

| Benchmark Category | Task Example | Result |
| :--- | :--- | :---: |
| **Python Algorithms** | QuickSort, palindrome verification, generators | ✅ **PASS** |
| **Tool Calling** | Structured JSON function argument extraction | ✅ **PASS** |
| **Arithmetic & Step Logic** | Multi-step word math reasoning | ✅ **PASS** |
| **Russian Dialogue** | Natural conversational flow without loop degradation | ✅ **PASS** |

---

## 🚀 Quickstart & Inference

### 1. In Ollama / LM Studio / llama.cpp (GGUF)

Pre-quantized GGUF binaries are provided directly in this repository:
- **`QuestionMark-Ultimate-0.6B-DUS32L-Q8_0.gguf`** (657 MB — Recommended: full precision, accurate EOS handling)
- **`QuestionMark-Ultimate-0.6B-DUS32L-Q4_K_M.gguf`** (480 MB — Ultra-compact footprint)

```bash
llama-cli -m QuestionMark-Ultimate-0.6B-DUS32L-Q8_0.gguf \
  -p "<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\nWrite a quicksort implementation in Python.<|im_end|>\n<|im_start|>assistant\n" \
  -n 512 --temp 0.2
```

### 2. In Apple MLX (Metal GPU)

```python
from mlx_lm import load, generate

model, tokenizer = load("KoarAI/QuestionMark-Ultimate-0.6B-DUS")
response = generate(
    model, 
    tokenizer, 
    prompt="<|im_start|>user\nHello! What tasks can you perform?<|im_end|>\n<|im_start|>assistant\n",
    max_tokens=256
)
print(response)
```

### 3. In Hugging Face Transformers (PyTorch / Safetensors)

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "KoarAI/QuestionMark-Ultimate-0.6B-DUS"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")

prompt = "<|im_start|>user\nWrite a Python function to check if a string is a palindrome.<|im_end|>\n<|im_start|>assistant\n"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256, temperature=0.2)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))
```

---

## 🛠️ Tool Calling Specification

The model natively adopts the standard ChatML tool-calling format:

**User Prompt:**
> *"What is the current weather in Krasnoyarsk?"*

**Assistant Output:**
```json
{
  "name": "get_current_weather",
  "arguments": {
    "city": "Красноярск"
  }
}
```

---

## ⚙️ Model Architecture Specifications

- **Base Architecture:** Qwen2.5 (Dense Decoder-Only Transformer)
- **Active Parameters:** ~600M (0.6B)
- **Layers:** **32** (Depth Up-Scaled via residual interpolation from 24)
- **Attention:** 14 query heads (2 KV heads, Grouped Query Attention)
- **Context Length:** 4096 tokens (Native 32K context window supported)
- **Training Pipeline:** TIES-Merging (Top 35% density trimming) + Safe On-Device ORPO RL

---

## 📜 Citation

If you incorporate this model, weights, or architecture in your work, please cite:

```bibtex
@misc{koarai2026questionmarkultimate,
  author = {KoarAI Studio},
  title = {QuestionMark-Ultimate-0.6B-DUS: On-Device 32-Layer LLM with Tool Calling and Cyrillic RL},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\\url{https://huggingface.co/KoarAI/QuestionMark-Ultimate-0.6B-DUS}}
}
```