Guide
Writing a book in Malayalam, Hindi or Tamil with AI
Models have got good at Indian languages, but not equally, and not in the way English users expect. The benchmarks, the token arithmetic, and the method that produces a book that reads like it was written in the language rather than translated into it.
How good are the models?
Better than two years ago, and still uneven. Google's IndicGenBench was the first broad generation benchmark across 29 Indic languages and found large gaps between Hindi, with abundant web text, and languages such as Malayalam and Odia. A 2025 benchmark, IndicParam, tested frontier models on knowledge questions in low-resource Indic languages and found the best of them scoring only around 45% on the hardest set. HinTel-AlignBench measured an 8-point drop from English to Hindi for the same tasks with the same models.
For book writing the practical picture is this: Hindi and Tamil prose from current large models is fluent and idiomatic most of the time. Malayalam, Kannada and Telugu are fluent but drift toward a formal, newspaper register and occasionally into unnatural calques of English sentence structure. Bengali sits in between. Smaller and older models are noticeably worse in every Indic language; the choice of model matters more than the prompt.
Why a Malayalam chapter costs three times an English one
Model pricing is per token, and tokenisers were built on English. A Malayalam word in its script often splits into six or seven tokens, a Hindi word into four or five, an English word into one or two. A 2,500-word Malayalam chapter can consume 17,000 tokens where the English equivalent uses 5,500. That has two consequences authors notice:
- Cost. The same book costs roughly three times as much to generate in Malayalam as in English. Neubook Write's estimates account for this by language, so the price you see before planning is the price you pay.
- Cut-off chapters. Tools that budget output by a fixed token count will stop a Malayalam chapter in the middle of a sentence. The budget has to scale with the script.
Research is chipping at this with Indic-specific models and tokenisers; see Paramanu and the survey of models for low-resource languages for the state of the field.
Write in the language; do not translate into it
The single biggest quality difference comes from asking for the book in Malayalam from the start, with the outline, examples and interview answers in Malayalam, rather than generating English and translating. Translated books carry English rhythm: long subordinate clauses, "not only... but also", the adjective triads. Books written natively in the language use its own devices, such as the Malayalam habit of ending a paragraph on a short verb-final sentence, or Hindi's proverbs.
A method that works
- Answer the interview in the target language, including place names, kinship terms and food. These anchor register.
- Choose the register explicitly. Spoken Malayalam and literary Malayalam differ more than spoken and written English do. Tell the tool which one, and whether English loanwords are welcome (they usually are in nonfiction, and usually not in devotionals).
- Give a voice sample in the language, even a WhatsApp message or a diary page. It fixes the model's tendency toward newspaper prose.
- Use scripture in the original translation your readers use. A Malayalam Bible devotional should quote the Sathyavedapusthakam, not a machine rendering of the NIV.
- Have a native reader check one chapter early, before all of them are written. Most problems are systematic and fixable in the style instructions.
- Do not run English AI detectors on it. They are meaningless on Indic text; see our note on detectors.
Where the readers are
Amazon.in sells Kindle books in Hindi, Tamil, Malayalam, Marathi, Gujarati and more; Google Play Books and Pratilipi carry regional-language readers that English-language charts never show. Print-on-demand is still limited for Indic scripts, so for paperbacks the usual route is a local printer with a PDF interior, which is why Neubook Write's KDP paperback export also produces a plain print-ready PDF.
Malayalam, Hindi, Tamil and more, written natively. Neubook Write budgets tokens by script, routes each language to the model that writes it best, and asks its two questions in your language.
Start a book, freeSources
- IndicGenBench: A multilingual benchmark to evaluate generation capabilities of LLMs on Indic languages
- IndicParam: Benchmark to evaluate LLMs on low-resource Indic languages
- HinTel-AlignBench: Hindi-Telugu benchmark with English-aligned samples
- Paramanu: Compact monolingual language models for low-resource Indian languages
- Large multimodal models for low-resource languages: A survey
- First Workshop on Language Models for Low-Resource Languages (LoResLM 2025)