--- language: - ne - en license: apache-2.0 base_model: Qwen/Qwen2.5-7B-Instruct pipeline_tag: text-generation tags: - sanu-ai - nepal - nepali - gguf - ollama - llama-cpp - quantized - bilingual - south-asia - low-resource ---
# SANU AI 7B v0.2 GGUF ### Run Nepal's First AI Locally — 3 Commands, 5 Minutes **Ollama Ready | Q4_K_M (4.68 GB) | Q8_0 (8.10 GB)** [![License](https://img.shields.io/badge/License-Apache_2.0-blue.svg)](https://opensource.org/licenses/Apache-2.0) [![LoRA](https://img.shields.io/badge/LoRA-Adapter-purple.svg)](https://huggingface.co/Haubaa/SANU-AI-7B-v0.2) [![Ollama](https://img.shields.io/badge/Ollama-Ready-green.svg)](#quick-start) [![Made in Nepal](https://img.shields.io/badge/Made_in-Nepal_%F0%9F%87%B3%F0%9F%87%B5-red.svg)](#) --- *From zero to chatting with Nepal's AI in under 5 minutes.*
--- ## Quick Start (Copy-Paste Ready) ### Step 1: Install Ollama ```bash # macOS / Linux curl -fsSL https://ollama.com/install.sh | sh # Windows — download from https://ollama.com/download ``` ### Step 2: Download SANU AI ```bash # Download the Q4_K_M model (4.68 GB — recommended for most users) curl -L -o sanu.gguf https://huggingface.co/Haubaa/SANU-AI-7B-v0.2-GGUF/resolve/main/sanu-ai-v02-q4_k_m.gguf ``` ### Step 3: Create & Run ```bash cat > Modelfile << 'EOF' FROM ./sanu.gguf SYSTEM "You are SANU AI, Nepal's first agentic AI assistant. You speak Nepali and English fluently. You have deep knowledge of Nepal's history, culture, geography, and current affairs." PARAMETER temperature 0.7 PARAMETER num_ctx 2048 TEMPLATE "<|im_start|>system\n{{ .System }}<|im_end|>\n<|im_start|>user\n{{ .Prompt }}<|im_end|>\n<|im_start|>assistant\n" EOF ollama create sanu -f Modelfile ollama run sanu ``` **That's it. You're now chatting with Nepal's first AI.** --- ## Available Models | Model | Size | RAM Required | Speed | Best For | |-------|------|-------------|-------|----------| | **Q4_K_M** (Recommended) | **4.68 GB** | 8 GB+ | Fast | Daily use, laptops, most GPUs | | Q8_0 (High Quality) | 8.10 GB | 12 GB+ | Moderate | Maximum quality, research | --- ## What Can SANU Do? ### Talk in Nepali ``` >>> नमस्ते, कस्तो छ? नमस्ते! मलाई नेपालका सबैभन्दा प्रशंसित AI सहयोजक SANU मा बुझाएका हुँ। ``` ### Nepal Knowledge ``` >>> Tell me about Nepal Nepal, officially the Federal Democratic Republic of Nepal, is a landlocked country in South Asia bordered by India to the south and east, China (Tibet) to the north... Kathmandu is its capital and largest city. The Himalayas form part of Nepal's northern border and include Mount Everest, the world's highest peak... ``` ### Write Code ``` >>> Write a Python hello world that prints namaste in Nepali **THINK:** To write a Python script that prints "Namaste" in Nepali... ```python def print_namaste(): name = u'नमस्ते' print(name) print_namaste() ``` ### Answer in Pure Nepali ``` >>> नेपालमा कति जिल्ला छन्? नेपालको सङ्घीय शासन प्रणालीले ७५ जिल्लाहरू नियुक्त गर्दछ... ``` --- ## v0.1 → v0.2: What Changed?
What v0.1 v0.2 Improvement
Training Data 290 samples 10,096 samples 35x more data
Loss 1.3724 0.795 42% lower
Hindi Contamination Severe Minimal Clean Nepali
Identity Confused (says "Qwen") "I am SANU AI" Fixed
Nepal Knowledge Surface-level Deep (Wikipedia + News) Massive upgrade
LoRA Rank r=16 r=32 2x model capacity
Q4_K_M Size 4.36 GB 4.68 GB Same ballpark
Cost $0 $0 Still free
### Response Comparison: v0.1 vs v0.2 | Prompt | v0.1 | v0.2 | |--------|------|------| | तिमी को हो? | "मैं Qwen हूँ, Alibaba Cloud द्वारा बनाया गया..." (Hindi, wrong identity) | "म SANU AI हुँ, नेपालको पहिलो AI सहायक" (Nepali, correct) | | Tell me about Nepal | Generic 2-3 sentences | Detailed: history, geography, politics, culture, economy | | नमस्ते, कस्तो छ? | "नमस्ते! मैं ठीक हूँ, धन्यवाद!" (Hindi mixed) | Pure Nepali response with SANU identity | | Write Python code | Basic output | Structured THINK/PLAN/CODE/VERIFY format | | सगरमाथा बारेमा | "माउंट एवरेस्ट दुनिया का सबसे ऊंचा पर्वत है" (Hindi) | Detailed Nepali: coordinates, history, UNESCO heritage | --- ## Training Details | Spec | Value | |------|-------| | Base Model | Qwen 2.5 7B Instruct | | Method | QLoRA 4-bit | | LoRA Rank | r=32, alpha=64 | | Data | 10,096 Nepali samples (Wikipedia + News + Custom) | | Loss | 0.795 | | GPU | NVIDIA P100 (Kaggle free tier) | | Time | ~6 hours | | Cost | **$0** | --- ## System Requirements | Component | Minimum | Recommended | |-----------|---------|-------------| | **RAM** | 8 GB | 16 GB | | **GPU VRAM** | Not required | 6 GB+ (for GPU acceleration) | | **Disk** | 5 GB (Q4) / 9 GB (Q8) | 10 GB | | **OS** | Windows 10+ / macOS 12+ / Linux | Any modern OS | --- ## Roadmap | Version | Data | Status | |---------|------|--------| | v0.1 | 290 samples | Released | | **v0.2** | **10,096 samples** | **Current** | | v0.3 | 50K+ samples | Coming Soon | | v1.0 | 200K+ samples | In Planning | --- ## Why This Matters **30 million Nepali speakers. 125+ languages in Nepal. Zero dedicated AI models — until now.** Every global LLM treats Nepali as noise. They confuse it with Hindi. They know nothing about Dashain, dal bhat, or the Malla dynasty. SANU AI is different — trained on real Nepali text from real Nepali sources. And it cost **$0** to build. --- ## Citation ```bibtex @misc{sanu-ai-2026, title={SANU AI: Nepal's First Agentic AI Foundation Model}, author={Haubaa}, year={2026}, publisher={HuggingFace}, url={https://huggingface.co/Haubaa/SANU-AI-7B-v0.2-GGUF} } ``` ---
**Built in Nepal 🇳🇵 | For Nepal | For the World** *Training cost: $0 | Dreams: Priceless*