Research

Models for
India’s languages.

AI models charge by the token, and Indian languages need far more tokens than English for the same sentence. We measure that gap, build tokenizers that close it, and train our own models on top.

Paper 1Preprint, November 2026

The Indic Token Tax in 2026

The same sentence costs more to send to an AI model in an Indian language than in English. We measured how much.

Nine public tokenizers, all 25 languages of India in FLORES-200
On GPT-4o: 1.57x the tokens of English for Hindi, 5.01x for Odia
Code, pinned artifacts and raw token counts are public

Code and data ↗

Paper 2Preprint coming soon

A tokenizer that needs fewer tokens than GPT-4o, including on English

Trained from scratch on human-written text in about 200 languages.

At GPT-4o’s vocabulary size: 6.1% fewer tokens on English
Fewer tokens on all 25 languages of India (median 0.74x GPT-4o’s count)
Fewer tokens on 161 of 204 languages

Tokenizers and results ↗

ModelIn training

Havon LM

Our own language models, built from scratch, starting with Marathi, Hindi and English.

124M parameters, trained on 2.5B tokens
Writes fluent Marathi, Hindi and Hinglish
Still a small base model: it invents facts. Next: a 1B model

Open by default

The code, the tokenizers and the raw token counts behind our papers are public, so anyone can check the numbers. github.com/HavonAI/indic-token-tax ↗

Working on Indian-language AI, or want to collaborate? hello@havon.ai