The Navdyut model family

Assamese intelligence, end to end.

Language, documents, and voice, built as one connected stack on a shared Assamese foundation. Open research proves the work. Proprietary models carry it into institutions.

Explore the system →
Abstract woven pathways moving through layers of a language model
One foundation · three modalities
One connected system

Built together, not bolted together.

Eri opens the research. The tokenizer holds the script. Muga supplies intelligence. Lipi and Bani carry that core into documents and voice.

01
Language

Assamese that begins with Assamese.

Script, corpus, and language model designed as one line.

Infrastructure

Tokenizer

Keeps Assamese text in meaningful pieces instead of fragmenting script. Ships with Eri; Muga uses the same vocabulary.

32K vocabularyScript-aware
Institutional

Muga

Flagship Assamese language model for self-hosted systems, built on the shared 32K tokenizer.

30B parametersSelf-hostable
Open research

Eri

Foundation model pretrained from scratch and released openly: proof of our training journey, including a 240M-parameter function-calling release on Hugging Face.

240M parametersOpen weightsHugging Face
02
Documents

Turn archives into usable knowledge.

Vision systems built for Assamese print and handwriting.

Extracts Assamese from books, forms, gazettes, and newspapers.

Vision → textPrinted script
Handwriting

Lipi Hand

Reads land records, correspondence, and manuscripts.

Vision → textHandwriting
03
Voice

For people who should never have to type.

Speech tuned for Assamese dialects and phonology.

Speech → text

Bani STT

Holds up across accent and dialect variation.

Dialect-aware
Text → speech

Bani TTS

Tuned for native Assamese phonology and sound distinctions.

Native phonology
Voice → voice

Bani

Bani STT, Muga, and Bani TTS joined into one spoken interface.

End to end
Research depth

Open where it helps. Proprietary where it counts.

Foundational models trained up to 1.5B parameters at 2.7× Chinchilla-optimal token count. That means more training data per parameter, for efficient and rigorous learning. Commercial deployments stay closed.

Hugging Face ↗ Work with us →