eulogik/Bharat-Tiny-LLM-v3

Hugging Face
TEXT GENERATIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:2BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 26, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Warm

Bharat-Tiny-LLM v3 by eulogik is a 1.7-billion-parameter small language model (SLM) based on Qwen3-1.7B-Base, designed for Hindi, Hinglish, and English. It features injected Devanagari subword tokens, improving Hindi text compression by 36.8% and achieving over 56% on GSM8K-Hindi math problems. This model is optimized for fully offline operation on resource-constrained edge devices, including phones, laptops, and browsers, with a Q4_K_M GGUF size of 1.0 GB.

Loading preview...

Bharat-Tiny-LLM v3: Bilingual Edge LLM

Bharat-Tiny-LLM v3, developed by eulogik, is a 1.7-billion-parameter small language model (SLM) specifically engineered for Hindi, Hinglish, and English on resource-constrained devices. Built upon Qwen3-1.7B-Base, it incorporates 300 injected Devanagari subword tokens, significantly enhancing its performance in Indian languages.

Key Capabilities & Features

  • Bilingual Proficiency: Understands and generates natural Hindi, Hinglish, and English.
  • Efficient Hindi Processing: Achieves 36.8% better Hindi text compression than the base tokenizer, with an 18.0% saving for mixed Hinglish+Hindi.
  • Mathematical Reasoning: Solves over 56% of grade-school math problems in Hindi (GSM8K-Hindi), a substantial improvement from previous versions.
  • Edge Device Optimization: Runs fully offline on devices as small as 1.0 GB (Q4_K_M GGUF), compatible with llama.cpp, Ollama, and browser WebGPU.
  • Privacy-focused: Operates entirely client-side, ensuring no data leaves the device.
  • Commercial Use: Licensed under Apache 2.0, allowing for free commercial use.

Ideal Use Cases

  • Offline AI Assistants: Perfect for applications requiring an offline Hindi LLM on mobile or embedded devices.
  • Bilingual Chatbots: Suitable for conversational agents needing to interact in Hindi, Hinglish, and English.
  • Educational Tools: Can be used for math problem-solving in Hindi, particularly for grade-school level tasks.
  • Resource-Constrained Environments: Designed for deployment on devices with limited computational power and memory.

Finetunes
3 models