Bharat-Tiny-LLM
v3Apache 2.0Frontier Vernacular Edge LLM Powered by Brahmi Tokenizer
ABOUT THIS MODEL
Edge-first Hindi/Hinglish foundation model powered by our proprietary Brahmi Tokenizer. Delivers 33–45% fewer tokens on Indic scripts with zero byte-fragmentation. Supports GGUF, WebGPU, llama.cpp, and transformers.js. Tool-use capable with GSM8K Hindi math reasoning. 100% on-device and air-gapped.
KEY HIGHLIGHTS
- ✓Powered by Eulogik Brahmi Tokenizer: eliminates Devanagari byte fragmentation
- ✓33–45% token compression over standard LLaMA/GPT tokenizers on Indic text
- ✓Full offline sovereignty: runs via WebGPU in browser and llama.cpp on edge CPU
- ✓Purpose-engineered for vernacular banking, legal, and citizen service workflows
SPECIFICATIONS
TokenizerBrahmi Tokenizer (Eulogik)
Token Efficiency33–45% fewer tokens
Base ModelQwen3-1.7B
LanguagesHindi · Hinglish · English
FormatsGGUF · WebGPU · SafeTensors
Deploymentllama.cpp · In-browser · CPU
FeaturesTool-Use · Math (GSM8K) · RAG
LicenseApache 2.0
Downloads377+
QUICK START
Get started in minutes
Install
pip install transformersUsage
from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained("eulogik/Bharat-Tiny-LLM-v3")
model = AutoModelForCausalLM.from_pretrained("eulogik/Bharat-Tiny-LLM-v3")
# Native Brahmi tokenization preserves morphology with zero fragmentation
inputs = tokenizer("नमस्ते! आज का मौसम कैसा है?", return_tensors="pt")
output = model.generate(**inputs, max_new_tokens=200)
print(tokenizer.decode(output[0], skip_special_tokens=True))Need help deploying Bharat-Tiny-LLM?
We offer consulting, custom fine-tuning, and on-premise deployment for all our models.