The-AI-Alliance/tapestry
Project Tapestry aims to give every nation and participant frontier AI they can call their own — uniting a global consortium to train a shared frontier model family from which partners build and own sovereign models aligned to their national, socio-cultural, and industrial needs.
What it solves
Most large AI models are trained on data that is not very diverse and are controlled by a few powerful organizations. Project Tapestry aims to change that by building a new family of foundation models trained on a much larger and more varied set of data. The goal is to enable "sovereign AI" — each partner keeps ownership of their data and compute, and can build their own derivative models based on the consortium-trained base models.
How it works
Tapestry is a consortium of organizations and individuals that bring together expertise, datasets, and compute resources. The repository contains the code and technical documentation for the project's open-source training platform. The code is organized into three subsystems: data (governance and management), training (distributed training and tuning), and infrastructure (underlying infrastructure). There is a prototype for consortium training (similar to federated learning) and a staging area (contrib/) where contributors can propose proofs-of-concept, experiments, and new ideas. The project also has active work streams on cultural alignment (using the Inglehart–Welzel Cultural Map as a metric), consortium training methods, and responsible data use.
Who it’s for
Tapestry is for organizations and individuals who want to participate in building sovereign AI. This includes partners contributing data, compute, or funding, as well as developers and researchers interested in large-scale training, federated learning, data governance, and cultural alignment. Contributors can join work groups, propose ideas via the contrib/ directory, or help define requirements for the first work streams.
Highlights
- Global consortium approach to create a foundation model family on a more diverse corpus than ever before.
- Open-source training platform with code for data governance, distributed training, and infrastructure.
- Focus on cultural alignment and re-alignment, initially using the Inglehart–Welzel Cultural Map.
- Prototype for consortium training that balances centralized and distributed methods (federated learning).
- Lightweight contribution process: new ideas go into
contrib/as proofs of concept via pull requests. - Clear governance and technical documentation under
docs/, with work groups forming in several areas.
Related
- Project
- Project
- Project
- Project