Project Tapestry: frontier AI, built together
A global effort with the AI Alliance to build AI models no single country could build alone, while each partner keeps its own data and models. We are a founding contributor.
A third way to build frontier AI
A country that wants capable AI usually has two choices: depend on a few large labs, with data it cannot inspect, or build alone, which few can afford and which rarely reaches the frontier.
Tapestry is a third way. Partners pool data, compute and talent to train a shared base model, then build their own versions on top, for their own languages, laws and needs.
Share what the model learns, not the data
Federated training: the data stays where it is, and only what the model learns from it travels.
A shared base model
Trained by the consortium and sent to every partner.
Train at home
Each partner trains it on its own data, computers and rules.
Send updates, not data
Only model updates travel back, to be reviewed and merged into the next base.
Adapt locally
Fine-tuning, safety and deployment stay in each partner's hands.
The first proof: India and Australia, one model
Completed on 1 September 2026, it tested whether organisations in different countries could train one model together. We led it, with Monash University in Australia.
No training data crossed borders
Our node in India and Monash's in Australia each tuned the same 7 billion parameter model on their own cultural data, and exchanged only compressed model updates.
Training kept going through dropouts
When the link dropped, each side kept training on its own, waiting no longer than a set delay.
Merging more often gave a better model
The more often the nodes merged their updates, the better the model fit both cultures. Fewer merges scored about a third worse.
Built on SLAKSHNA
Our side ran on SLAKSHNA, the federated training framework built by BharatGen's Delhi node at IIIT Delhi.
Milestone One runs from September to November 2026, on consortium training, cultural alignment and data governance. Read the Milestone Zero findings ↗
What we bring to Tapestry
A founding contributor
We joined in June 2026, with the support of the IndiaAI Mission, and anchor India's role.
Co‑leading distributed training
India co-leads Tapestry's distributed training, drawing on how we train across our own consortium.
Proof that it works at country scale
Text, speech and vision AI models built from scratch for India, when funding, compute, data and institutions come together.
"Welcome to Project Tapestry, BharatGen!"
"India has long believed that science advances fastest when it is open and shared."
From New Delhi to Milestone Zero
- The idea takes shape in New DelhiAt the India AI Impact Summit, the AI Alliance asks how the open-model community can work together at a scale that matters.
- First Tapestry workshop, ParisAround 30 AI researchers and leaders plan the architecture and roadmap. Prof. Ganesh Ramakrishnan opens the second day with BharatGen's story.
- We join as a founding contributorBharatGen commits to Tapestry in the week of the G7 Summit, with India co-leading distributed training.
- Milestone Zero completeOur proof of concept with Monash trains one model across India and Australia without moving any training data.
Follow Project Tapestry
Read the AI Alliance's reports on the work so far, or talk to us about taking part from India.