MODEL · ON-DEVICE EDGE
Private preview · GA Sep 2026Tatva
तत्त्वSovereignty that fits on the device.
Tatva is AgentAnywhere's on-device edge model — small, efficient, and built to run offline, air-gapped, and on modest hardware where connectivity and cloud are not options. It is also the family we train from scratch in India, from token one, on our own Indic corpus. In private preview now.
- Runs
- Offline and air-gapped, on modest hardware — no connectivity required.
- Lineage
- Trained from scratch in India, from token one, on our own Indic corpus.
- Data
- Never leaves the device — the strongest form of sovereignty there is.
- Languages
- Indic languages and English, built in from the ground up.
The perimeter isn't always a data centre.
Some of the places that most need intelligence have no cloud to call: a branch on an unreliable link, a field device in a remote district, an air-gapped facility where connectivity is forbidden by design. For those, “runs in your perimeter” has to mean “runs on this box, disconnected” — not “runs in your region.”
Tatva is built for that edge. It is small and efficient enough to run on-device, offline, with the model and the data never leaving the hardware in your hand. It is the same sovereignty principle as the rest of the platform, taken to its limit.
What Tatva is built for.
Offline & air-gapped
Runs with no network and no cloud dependency — for facilities and workflows where connectivity is unavailable or not permitted.
Modest hardware
Small and efficient by design, so it runs on the kind of hardware you can put at the edge — not a rack of accelerators.
The field and the frontier
For branches on thin links, remote and rural deployments, and disconnected environments where the work still has to get done.
Built from scratch in India
Tatva is our own from-scratch lineage — trained from token one on an Indic corpus we cleaned and sealed ourselves, not a fine-tune of someone else's weights.
Indic from the ground up
Indian languages and English are in the training mix by design, controlled and measured — not inherited by accident from a scrape.
Provenance, even at the edge
The same receipts discipline as the rest of the platform: what the model is, and what it learned from, is a sealed, auditable record.
Sovereign to the last inch.
With Tatva, the data never leaves the device — there is no perimeter to breach because nothing is transmitted. For defence-adjacent, critical-infrastructure, and privacy-critical field use, that is the strongest guarantee on offer.
And because Tatva is trained from scratch on a corpus we build and seal ourselves, its lineage is our own: a model whose composition you can be shown, not a black box borrowed from elsewhere. The training corpus is cleaned and sealed by our data factory, with a build manifest every run cites.
Who Tatva is for.
Field operations, remote and rural deployments, on-prem appliances, and air-gapped facilities — anywhere the work happens at the edge of, or entirely outside, the network, and the data must stay on the device.
Where Tatva is today — stated plainly.
Tatva is in private preview, with general availability targeted for September 2026. The from-scratch training programme behind it is underway and measured; where we cite a figure elsewhere on this site, the coverage label — measured, sampled, or estimated — is part of the claim.
We are taking design partners with real edge and offline requirements. Talk to us if your deployment lives where the cloud does not reach.
FAQ
Frequently asked questions.
- What is Tatva?
- Tatva is AgentAnywhere's on-device edge model — small and efficient, built to run offline and air-gapped on modest hardware where cloud and connectivity are not options. It is also the family we train from scratch in India, from token one, on our own Indic corpus. It is in private preview, with general availability targeted for September 2026.
- Does Tatva need an internet connection?
- No. Tatva is built to run offline and air-gapped, on-device, with the model and data never leaving the hardware. It is designed for branches on thin links, remote and rural deployments, and facilities where connectivity is not permitted.
- Is Tatva a fine-tune of another model?
- No. Tatva is our own from-scratch lineage — trained from token one on an Indic corpus we clean and seal ourselves, with a build manifest every training run cites. Its composition is an auditable record, not a borrowed black box.
- Can we pilot Tatva at the edge now?
- Tatva is in private preview and we are onboarding design partners with real offline and edge requirements. Talk to us about your deployment; general availability is targeted for September 2026.
Run sovereign intelligence where the cloud can't reach.
Tatva is in private preview with design partners, GA targeted September 2026. If your deployment is offline, air-gapped, or at the edge, talk to us.