🗣️ Bharat-Tiny-LLM v3
Bilingual Hindi / Hinglish / English edge LLM — runs 100% in your browser. Private, offline, free. build 2026-09-16h
-blue){kind=icon}
Loading model (first run downloads ~1.6 GB, then cached)...
Why Bharat-Tiny-LLM v3?
| Feature | Benefit |
|---|---|
| 1.7B params, Q4 (~1 GB) | Runs on phones, laptops, edge devices |
| Devanagari token injection | 36.8% better Hindi token compression |
| SFT on Hindi math + QA | 56%+ on GSM8K-Hindi (vs 0% in v2) |
| 26.6% fewer bits/char | Stronger Hindi language modeling |
| In-browser WebGPU | No server, no data leaves your device |
Run it yourself
llama.cpp: llama-cli -m bharat-tiny-llm-v3-q4_k_m.gguf -p "नमस्ते"
Ollama: ollama create bharat-v3 -f Modelfile
Python: pip install llama-cpp-python then load eulogik/Bharat-Tiny-LLM-v3
FAQ
Is my data private? Yes — inference runs entirely in your browser via WebGPU. Nothing is sent to a server.
Which languages? Hindi (Devanagari), Hinglish (Romanized Hindi+English), and English.
Does it need internet after first load? Only the first load downloads the model; afterwards it's cached offline.
What can it do well? Conversational Hindi/Hinglish, factual Q&A, and grade-school math word problems in Hindi.