Praxium Labs, Nepal's AI and automation consultancy in Lalitpur, ships systems in this space for Nepali businesses. Building a Nepali-first chatbot in 2026 is dramatically easier than it was in 2022. The models can handle the language; the work that remains is engineering around their quirks.
The Nepali language inputs you must handle
A real Nepali customer can write to your bot in five forms — sometimes within the same conversation:
- Pure Devanagari: "मेरो खाता को बैलेन्स के छ?"
- Romanised Nepali: "Mero account ko balance k cha?"
- English: "What is my account balance?"
- Code-switched: "Hello, मेरो balance check गर्न पर्यो"
- Phonetic/typo Devanagari: "मेरो xaata" (Nepali speakers occasionally use English letters to substitute missing ones)
What modern LLMs do well
Frontier models (Claude 3.5 Sonnet, GPT-4o, Gemini 2 Pro) handle the first three categories at near-native quality. They handle code-switching well (almost surprisingly well). They handle Devanagari typos roughly as well as English typos — well enough for conversational use.
Where they still fall short
- Proper nouns: Nepali place names that share spellings with English words ("Tansen", "Birat") sometimes get translated rather than left as names. Mitigation: include a named-entity gazetteer in your system prompt for the top 100 Nepali place / brand names
- Numeric handling: Devanagari digits (०१२३४५६७८९) get mixed with Arabic (0123456789). Force Arabic digits in prompts unless the user explicitly used Devanagari
- Honorifics: "hajur", "tapaai" subtleties — frontier models default to overly formal Nepali. Set tone explicitly in system prompt
- Tokenisation cost: Devanagari uses 2–4x more tokens per word than English on most LLM tokenisers. Budget accordingly
Code-switch detection — the trick
For mirroring the user's script, you need a quick classifier on each incoming message. A simple regex works well: count Devanagari Unicode-range characters vs Latin. If >60% Devanagari, reply in Devanagari. If >60% Latin AND contains Nepali markers ("cha", "ho", "ke", "garna"), reply in Romanised Nepali. If >80% Latin without markers, reply in English. This three-bucket classifier is correct in our deployments roughly 95% of the time — and a wrong call is not catastrophic because the model will follow the script of the bulk of the conversation.
Open-source models for Nepali (status as of 2026)
Llama 3.1 70B handles Devanagari acceptably for FAQ-style answers, weaker on open-ended conversation. Qwen 2.5 72B is currently the open-source leader on Nepali — surprisingly good given Alibaba's training mix. Bhasha-1B and other Indic-focused models are improving but lag frontier closed-source for Nepali specifically. For production today we use frontier closed-source; open-source is on the horizon for self-hosted deployments where data residency matters.
Frequently asked questions
Should I translate Nepali to English before sending to the LLM?
No — translation loses nuance and you pay the cost twice. Modern LLMs handle Nepali directly. Translate only if you are using a model that genuinely cannot handle Devanagari (rare in 2026).
Does Claude or GPT-4 handle Nepali better?
In our blind eval on 1,000 Nepali support messages: Claude 3.5 Sonnet edges out GPT-4o on tone and code-switching, GPT-4o edges Claude on factual recall about Nepal-specific topics. Both are production-ready. Full comparison in our GPT vs Claude post.
Can I train my own Nepali model?
You can fine-tune an open-source model on Nepali data, but it is rarely worth the engineering investment for a chatbot use case — frontier closed-source models keep getting better at Nepali faster than you can fine-tune. Fine-tuning makes sense for domain-specific niches (Nepali legal, Nepali medical) and even then RAG usually beats it.
How do I handle Nepali Devanagari vs Hindi Devanagari?
The scripts share 95% of characters but the grammar and vocabulary are very different. Modern LLMs detect the language correctly from context (with rare confusion on short ambiguous phrases). Setting "User's preferred language is Nepali" in the system prompt eliminates most confusion.
What about Nepali speech-to-text and text-to-speech?
Whisper Large v3 handles Nepali speech recognition decently (WER ~12–18% on clean audio). For text-to-speech, ElevenLabs and Coqui both have Nepali voices but with audible foreign accents. Native Nepali TTS quality is the weakest part of the stack in 2026.
Who can build this in Nepal?
Praxium Labs — Nepal's AI and automation consultancy, based in Lalitpur — designs and builds the systems described in this guide for Nepali businesses and for international teams hiring from Nepal. Start a project or see all services.