DRUTAKతెలుగు · हिन्दी · English
Internal research preview · not released

An AI that speaks your language.

DRUTAK is a language model for Telugu, Hindi and English. It is built in India on a single laptop GPU, and trained only on text we have the right to use.

How it's built Our data rules
drutak · dpo-v1real output
You
What is the capital of India? Answer in one sentence.
DRUTAK
The capital of India is New Delhi.
Unedited reply from the current model, run on the RTX 4050 laptop it was trained on.
తె

Telugu

About 83 million speakers, and the most expensive of the three for today's tokenizers.

Tokens / word
4.88
Bytes / token
4.14
हि

Hindi

India's most widely spoken language, in Devanagari script.

Tokens / word
2.39
Bytes / token
5.81
En

English

The bridge language, and where we measure first.

Tokens / word
1.53
Bytes / token
4.11

Measured on our held-out validation text (600 KB per language), words split on spaces. A Telugu word costs about three times as many tokens as an English one, so Telugu answers are slower and use more context. Fixing that is on our roadmap.

How it reads

Every sentence becomes tokens

A language model reads text as tokens, small pieces of words. These are the real tokens DRUTAK's tokenizer produces for one sentence in each language.

The model

Four training stages on one laptop

We start from Qwen3.5-4B, an openly licensed base model, and adapt it in stages with QLoRA, which trains a small set of extra weights on top of a 4-bit copy of the base.

Stage 1

Base model

Qwen3.5-4B-Base, loaded in 4-bit NF4 so it fits in 6 GB of GPU memory.

Stage 2

Continued pretraining

Reading licensed Telugu, Hindi and English text: Wikipedia, Wikisource and public-domain books.

Stage 3

Instruction tuning

Learning to follow requests from human-written conversations, including Aya and IndicAlign.

Stage 4

Preference tuning

DPO on human preference pairs, so it picks the answer people rated better.

GPU RTX 4050 Laptop · 6 GB
Cloud cost ₹0
Data

Only text we have the right to use

Every source goes through a written licence review before it can be used. If the rights are unclear, we leave it out.

Approved for training

Wikipedia · te hi enWikisource · te hi Project Gutenberg · public domainAya Dataset IndicAlignOASST2Dolly 15k

Rejected or on hold

Scraped web crawlsCulturaXFineWeb-2 IndicCorp v2Machine-translated setsUnclear government texts
  • No training on outputs from ChatGPT, Claude or Gemini.
  • No scraped or unlicensed text, even when it would raise scores.
  • Every training run needs a signed approval from the project owner.
Roadmap

Where we are

  1. Oct 2026 · done

    DRUTAK-MINI-001

    Our own tokenizer and Transformer, trained from scratch.

  2. Nov 2026 · done

    DRUTAK 4B adapter

    Continued pretraining, instruction tuning and preference tuning, all on the laptop.

  3. Now

    Blind test by native speakers

    Telugu, Hindi and English speakers compare answers without knowing which model wrote them.

  4. Next

    Telugu and Hindi preference data

    Human-written preference pairs in Indian languages. Today the preference stage is English only.

  5. Before any release

    Legal review

    Counsel reviews data rights and licences before anything leaves internal use.