Indic Models Beyond Translation

The default failure of multilingual models on Indic languages starts at the tokenizer. Scripts like Devanagari, Telugu, and Tamil routinely consume 3–5× more tokens per unit of meaning than English, which means worse latency, worse cost, and effectively a shorter context window for the same task.

Fixing tokenization first changes every downstream number. A script-aware vocabulary with balanced allocation across families cuts token counts substantially and improves quality before a single additional training token is spent.

Data is the second problem. High-quality Indic corpora are scarce, code-mixed text is the norm rather than the exception, and transliteration into Latin script is extremely common in real user input. A model that only handles native script handles a minority of actual queries.

Evaluation is the third and most neglected. Translated English benchmarks measure translation artifacts. We build task suites natively in each language, sourced from real usage — government services, agriculture, health, and commerce — because those are the deployments that matter.