eulogik/Bharat-Tiny-LLM-v3
Bharat-Tiny-LLM v3 by eulogik is a 1.7-billion-parameter small language model (SLM) based on Qwen3-1.7B-Base, specifically designed for Hindi, Hinglish, and English. It features injected Devanagari subword tokens, improving Hindi text compression by 21.6% and achieving over 56% accuracy on GSM8K-Hindi math problems. Optimized for resource-constrained edge devices, it runs fully offline with a Q4_K_M GGUF size of 1.0 GB.
Loading preview...
Bharat-Tiny-LLM v3 Overview
Bharat-Tiny-LLM v3 is a 1.7-billion-parameter small language model (SLM) developed by eulogik, built upon the Qwen3-1.7B-Base architecture. It is specifically engineered for bilingual Hindi/Hinglish/English applications on resource-constrained edge devices.
Key Capabilities & Features
- Bilingual Proficiency: Understands and generates natural Hindi and Hinglish, alongside English.
- Enhanced Hindi Tokenization: Achieves 21.6% better compression for Hindi text compared to the base tokenizer, and 16.7% savings for mixed Hinglish+Hindi.
- Improved Hindi Reasoning: Solves over 56% of grade-school math problems in Hindi (GSM8K-Hindi), a significant improvement from previous versions.
- Edge Device Optimization: Designed to run efficiently on devices with limited resources, available as a 1.0 GB Q4_K_M GGUF file.
- Offline Operation: Functions entirely client-side, ensuring data privacy with no API calls or server interaction.
- Apache 2.0 Licensed: Free for commercial use, inheriting its license from the base Qwen3-1.7B-Base model.
When to Use This Model
Bharat-Tiny-LLM v3 is ideal for applications requiring a compact, efficient, and privacy-preserving language model for Indian languages. It is particularly well-suited for:
- Offline mobile or embedded applications needing Hindi, Hinglish, or English language processing.
- Educational tools focused on Hindi math word problems.
- Developers seeking an Apache-2.0 licensed model for commercial deployment on edge devices.