Kathmandu — In a quiet lab at Tribhuvan University's Kirtipur campus, a team of eight researchers has spent 18 months building something that could change how 30 million Nepali speakers interact with technology.
They've named it NepalaLM — a 7-billion parameter large language model trained on over 40GB of Nepali text including news, literature, legal documents, and social media. It runs entirely in Devanagari script and handles the grammatical complexity of Nepali — including its tense-heavy verb system — significantly better than any existing multilingual model.
NepaliWave sat with lead researcher Dr. Bikash Poudel for a hands-on demonstration. The model answered factual questions about Nepali law, wrote grammatically correct formal letters, and — most impressively — recognised and correctly handled the distinct Nepali used in the Terai versus the hills.
The implications go beyond convenience. Nepal's legal system, healthcare records, and government communications operate overwhelmingly in Nepali. An effective Nepali LLM could power court document analysis, medical record summarisation, and agricultural advisory systems that actually work for farmers who've never typed in English.
The team is releasing NepalaLM under an open licence next month. Dr. Poudel told NepaliWave: 'We don't want this locked behind a corporate API. This belongs to the Nepali people.'