TATVA NANO · TRAINED FROM SCRATCH IN INDIA
Tatva Nano
नैनोToken one was ours.
Tatva Nano is the smallest member of the Tatva family and the clearest proof of its lineage: a compact language model trained from zero in India, with no foreign base and no inherited weights. It writes India's languages in their own scripts, the Dravidian languages included, and runs fully offline on a laptop CPU. It is a capability proof; larger models are planned.
- Lineage
- From scratch. No base model, no inherited weights.
- Languages
- Writes Indian languages in their own scripts — Devanagari, Dravidian, Bengali.
- Runs
- Fully offline on a laptop-class CPU. No network, no GPU, no cloud.
- Scope
- A capability proof. Evaluations under way; no accuracy figure published yet.
Why a small model matters more than a big claim.
Almost every model sold as sovereign inherits its weights from somewhere. That is a reasonable engineering choice and an awkward regulatory one, because the first question a defence programme or a bank's model-risk team asks — what is it built on? — has to be answered with someone else's model.
Tatva Nano exists to make the other answer possible. It was initialised at random and trained by us, in India, with its own tokenizer: there is no base model underneath. Its training data is drawn from a number of sources, some of them openly available datasets from elsewhere, and we share the source list on request. It is small because a proof should be legible, not because small is the ambition.
What Tatva Nano is.
A from-scratch lineage
Random initialisation, our tokenizer, our run. No base model, so there are no inherited weights to explain in front of an auditor.
A tokenizer built for Indian scripts
Devanagari, the Dravidian scripts and Bengali are first-class in the vocabulary, not an afterthought bolted onto an English tokenizer.
Sources on record
The training sources are listed dataset by dataset, and we share the list on request.
Offline, on a laptop
The checkpoint runs on a laptop-class CPU with the network off. The demonstration is given that way on purpose.
What it does today, and what it does not.
A capability proof is only useful if its edges are drawn honestly.
Tatva Nano does
Continue a sentence in fluent script in a major Indian language — Hindi, Marathi, Bengali, Tamil, Telugu, Kannada, Malayalam among them — from a few words of prompt.
Run fully offline on ordinary hardware. Shuddhi screened and receipted most of its training text before training. The training sources are listed, and we share the list on request.
Serve as the checkpoint the rest of the from-scratch family grows from.
Tatva Nano does not
Act as an assistant or a knowledge base. It will invent facts and it will ramble; it is a language model, not an answer engine.
Carry a published accuracy figure yet. Evaluations are under way on a frozen set; the number appears when it is measured, not before.
Stand in for the larger checkpoints. Usefulness beyond the proof comes with the next rungs, which are planned on the same lineage.
Who asks for it, and why.
The buyers who care about Tatva Nano are the ones who get asked where a model came from.
Defence and strategic programmes
A model with no base model underneath — and that runs air-gapped, because it was never dependent on anything outside the box.
PSU banks and model-risk teams
Model risk management wants training-data provenance on record. With Tatva Nano the training sources are listed, and we share the list on request.
Government and public-sector programmes
Indic-first by construction, trained in India on Indian-language text, with no foreign base to disclose or to depend on.
See it run.
The demonstration does one honest thing. You give Tatva Nano the start of a sentence in any Indian language — a few words in Hindi, Tamil or Telugu — and it continues it in fluent script, on a laptop with the network switched off. It is a five-minute proof that the lineage is real, and we give it in person or on a call.
Talk to us about Tatva Nano.
Tatva Nano is a capability proof. Larger models are planned on the same from-scratch lineage and will be described here when they are measured, not before. If provenance is the question your reviewers ask first, this is the model to see.
Related
Where Tatva Nano fits.
FAQ
Frequently asked questions.
- What is Tatva Nano?
- Tatva Nano is the smallest member of AgentAnywhere's Tatva family: a compact language model trained from zero in India, with no base model and no inherited weights. It writes India's languages in their own scripts and runs fully offline on a laptop CPU. It is a capability proof.
- Is Tatva Nano a fine-tune of another model?
- No. Tatva Nano was initialised at random and trained by ShepHertz in India, with an Indic tokenizer built for Indian scripts. There is no base model and there are no inherited weights.
- Which languages can Tatva Nano write?
- Major Indian languages in their own scripts, including Hindi, Marathi and Bengali in Devanagari and Bengali script, and all four major Dravidian languages — Tamil, Telugu, Kannada and Malayalam. Given the start of a sentence, it continues it in fluent script.
- Can Tatva Nano answer questions or act as an assistant?
- Not today, and the page says so. Tatva Nano is a capability proof of a from-scratch lineage: it continues text fluently but will invent facts. Usefulness beyond the proof comes with the larger models planned on the same lineage.
- Does Tatva Nano have published accuracy numbers?
- Not yet. Evaluations are under way on a frozen evaluation set, and no figure is published until it has been measured there. What we can show you now is the list of training sources and a live offline demonstration.
Ask us where the model came from. Then watch it run offline.
Tatva Nano is a capability proof. If provenance is the first question your reviewers ask, we will show you the list of training sources and the checkpoint — on a laptop with the network off.