Every day, thousands of Nepalis engage with artificial intelligence in their work, asking to draft documents, translate, or inquire about public services. For an average administrative inquiry such as “How long does it take to renew my passport and how much does it cost?” the question in English would be parsed as 17 tokens by an AI. The same question in Devanagari 'मेरो राहदानी नवीकरण गर्न कति समय लाग्छ र शुल्क कति हो ?', 54 tokens by OpenAI’s GPT-4, 57 by Mistral, and 50 by Alibaba’s Qwen.
For a single user, this might just look like a minor technical glitch. But imagine the scale of the problem when our national digital services actually start integrating these AI models.
Take high-traffic platforms like the Nagarik App or Hello Sarkar. Hello Sarkar alone dealt with 69,724 citizen complaints last fiscal year, jumping 321% from the year before. Because they handle so much daily public engagement, if these systems rely on AI to process those queries, this token gap will immediately become a massive, invisible financial tax slapped directly onto our native alphabet. And that tax isn't going anywhere as long as current AI models and their underlying tokenizers remain so poorly optimized for Devanagari. The only real way forward is to demand and build technology that is actually designed to process our language efficiently.
To understand why, let us examine how tokens work.
An artificial intelligence parses natural language text into “tokens,” statistical chunks of characters, which allow the machine to understand it. Tokens serve as the atomic currency of modern AI. When using commercial APIs, users are charged based on tokens read/written. Furthermore, an AI model stores instructions in a “context window” whose capacity is also measured in tokens. Think of this window as the machine's short-term active memory. This becomes highly relevant when training or prompting an AI to perform tasks, because the heavy token cost of Devanagari quickly clutters this limited workspace. As a result, the system burns through most of its brainpower simply trying to read our Nepali commands, leaving very little room to actually execute them.
A pilot analysis of seven open-source and commercial large language models (LLMs) (OpenAI’s GPT-4 and GPT4o, Qwen, Mistral, Llama, DeepSeek, and Gemma) demonstrates that the “Devanagari Tax” is significant and systematic, particularly in older models. Testing ten citizen inquiries spanning passports, citizenship, road complaints, driving trials, and local taxes, Devanagari consistently cost 1.4 to 4.4 times more tokens than English for identical meanings.

The reason for this is down to engineering choices made during a model’s training. Under the ubiquitous UTF-8 encoding system, English letters use 1 byte per character while Devanagari uses 3 bytes. Since most internet content is in English, each English word gets its own token while Devanagari gets split into multiple tokens across characters. Essentially, the token-to-word ratio is model-dependent: English tends to be relatively token-efficient, while Nepali in Devanagari can require substantially more tokens for equivalent content, with the measured overhead varying considerably across tokenizers. Compared to a taxi meter which charges per distance driven, this is corresponding to a meter that charges per second instead of per kilometer. It is far more punishing for complex Devnagari vocabulary.
This computational bias creates an absurd incentive: it is almost always cheaper to abandon our script entirely and type Nepali in the Latin alphabet. Across most language models, Romanized Nepali slips through tokenizers far more easily than native Devanagari, effectively turning English lettering into an economic shelter.
Yet leaning on Romanized text exposes another layer of computational chaos. Because phonetic Nepali lacks standardized spelling, arbitrary typing habits alter the bill. A user dashing off an informal shortcut like “xaina” triggers a completely different token count than someone typing “chainaa” or “chhainaa.” While newer models with massive vocabularies such as GPT-4o have finally begun bridging this divide and bringing Devanagari close to parity with Latin script, the broader ecosystem has not caught up. Across older architectures and open-source models alike, native Devanagari remains an expensive liability, quietly penalizing anyone who chooses to write in their own alphabet.
This invisible tax has three implications: fiscal, cognitive, and sociolinguistic.
The primary issue is purely financial. Consider a public platform fielding a million monthly queries. If users submit brief, 19-token prompts using GPT-4o (priced at $2.50 per million input tokens as of September 2026), the English workload costs roughly $47.00. Processing those exact same queries in Devanagari pushes the bill to $67.75, which is a 44% markup or an extra $20.75 each month. Likewise, because output tokens carry a higher price tag than input tokens, this financial burden compounds even further the moment these AI models start generating their replies in Nepali.
The second is cognitive: memory management. A context window that holds 128,000 tokens can accommodate approximately 96,000 English words using the rough estimate of 1.33 tokens per word. However, the same token budget may represent substantially fewer words in Devanagari text with some older tokenizers, because tokenization efficiency varies across writing systems. As a result, an AI system may be able to retain less Nepali legal text within the same context window, potentially limiting how much of a long statute it can process at once.
The third consideration is sociolinguistic: Because Romanized Nepali is cheaper than Devanagari on most open and commercial models, developers have an economic incentive to create apps which nudge users to use Latin script (“mero nagarikta banauna k garnu parcha?”). This penalizes our own alphabet within our digital ecosystem.
These observations are consistent with research conducted at the Neural Information Processing Systems (NeurIPS) conference in 2023. Petrov et al. demonstrated that current tokenizers systematically charge higher prices (in terms of tokens) to non-English languages on the FLORES-200 benchmark. Specifically, they noted that per-token pricing transforms architectural bias into financial costs. However, this is an engineering choice and not an inevitability. India’s Sarvam-1 model demonstrated that increasing vocabulary size significantly reduces token counts for South Asian languages, particularly Hindi and Marathi languages which use the Devanagari script. More recent models such as Google’s Gemma-2 and GPT-4o demonstrate that reserving vocabulary slots for Devanagari decreases the “tax” by half. Locally, Himalaya AI announced intentions to build sovereign models (HimalayaGPT) alongside Devanagari-optimized tokenizers as part of their roadmap.
As Nepal’s National AI Policy 2082 promotes the development of state-owned computing infrastructure, it is crucial to recognize that software engineering precedes hardware decisions. Simply purchasing servers will not be sufficient to achieve digital sovereignty if the models which run on them implicitly hurt our linguistic and cultural heritage. Nepal must formalize requirements for tokenizer optimization in public bids, and fund local engineers to contribute to open-source tokenizer improvements before making billions of rupees of expenditure on cloud computing.
Nepal has already paid a premium for being landlocked and late to industrialize. In the age of artificial intelligence, we must ensure that we do not quietly pay an invisible tax on our own mother tongue.
The writer is a BSc CSIT graduate from Tribhuvan University and an AI/ML Engineer by practice working in the sphere of Generative AI, LLMs, RAGs, and multi-agent systems.






DR Congo’s Ebola Cases Top
India Shuts Down Iconic Delhi
False Hijacking Alert Triggers Major
Police Confiscate Robot That Groped
The Devanagari Tax: Why Artificial
eSewa Money Transfer Partners with