Project Tapestry Marks First Milestone, Expands Vietnam and India AI Ties

The AI Alliance announced on September 30, 2026, during United Nations General Assembly week in New York, that Project Tapestry had completed its first technical milestone and expanded sovereign AI collaborations with Vietnam and India.

Project Tapestry is building an open, global consortium in which nations and institutions collaborate on foundation models while retaining control of their data, sovereign models, and deployments. Under the architecture described on the project page, partner nodes keep sensitive data local, use it to compute weight updates, and share only the resulting model weights; any partner can build a fully owned sovereign derivative from the shared base model. The page lists Yann LeCun as the project’s chief science advisor and describes the AI Alliance as a nonprofit AI research and open-source technology coalition with more than 200 member organizations.

Vietnam Affirms Intent to Collaborate on Sovereign Model

During General Assembly week, AI Alliance and Project Tapestry leaders met in New York with Vietnam’s General Secretary and President Tô Lâm and senior government ministers. The discussions affirmed an intention to collaborate on a sovereign AI model aligned with Vietnam’s national AI strategy and priorities, including Vietnamese language, culture, knowledge, and national use cases. Vietnam has participated in Project Tapestry since its early stages, and the alliance described the expanded collaboration as a step toward turning the consortium model into national-scale sovereign AI development.

“Project Tapestry gives Viet Nam an opportunity to turn international collaboration into lasting national capability—developing AI that understands our language, culture, knowledge, and priorities, while building the engineers and institutions to operate, adapt, and advance the technology,” said Ambassador Hoang Anh Tuan, Consul General of Viet Nam in San Francisco.

Christopher Nguyen, chief architect of Project Tapestry, said Tô Lâm’s support for the direction underscored the project’s broader vision, pointing to parallel work across Viet Nam, India, Japan, and Bhutan.

India Briefing Precedes Mumbai Workshop

Consul General Binaya Srikanta Pradhan hosted Project Tapestry leaders at the Consulate General of India in New York during the week. Ganesh Ramakrishnan, founding director of BharatGen, Ritwik Banerjee of Stony Brook University, and Kaushik Bhatta, a partner and director at the AI Alliance, briefed the consul general on consortium-driven AI development, Tapestry’s approach, and Stony Brook’s recent addition to the consortium. Pradhan said AI built for India’s languages and cultures preserves agency.

India has been an early contributor to the project. BharatGen, the country’s government-backed, open-source foundation model stack, ran one of the milestone’s two proofs of concept under the guidance of Dr. Maneesh Singh, VP of machine learning, training jointly with Monash University across India and Australia while each institution’s data remained in its home country. The Gates Foundation named Ramakrishnan a 2026 Goalkeepers Champion during the same week, recognizing his leadership of BharatGen. BharatGen will host a Project Tapestry workshop in Mumbai on October 15, 2026, to plan the next phases of India’s participation.

Milestone Zero Findings

The alliance completed Milestone Zero on September 1, 2026, and published the findings on September 10, 2026. Two proofs of concept trained or tuned models across four sites in India, Australia, and two AWS regions, and eight further contributions arrived through the project’s open contribution process.

In the first proof of concept, the joint BharatGen and Monash University team tuned the OLMo 2 7B model locally at each node using disjoint partitions of the CultureInstruct dataset, periodically merging compressed LoRA weight deltas while exchanging no tuning data. Cultural alignment was measured with the GlobalOpinionQA evaluation using Jensen-Shannon Distance over seven rounds. The findings identify Round 3 as the best checkpoint on the equal two-region metric and caution that the preliminary results should not be interpreted as evidence that federation harmed performance.

The second proof of concept ran continued pre-training of OLMo 3 7B through the Flower federated framework across sites in Sydney and Virginia, coordinated from Ohio. According to the findings, the run trained on 12 billion tokens across the consortium using eight NVIDIA H100 GPUs per site, sustained approximately 44,400 tokens per second at each site, and completed two rounds with an independently verified final checkpoint, despite measured round-trip latency of 186 milliseconds and bandwidth of roughly 0.53 Gbit/s on the Ohio–Sydney route. The partitioned Dolma 3 training data never left the sites.

A separate study used LoRA fine-tuning on the Llama-3.2-3B-Instruct model and measured its distance to Vietnam’s position on the Inglehart-Welzel cultural map, moving from 2.46 to 1.35, which the team describes as 45% closer, while full MMLU moved from 63.2% to 62.4%, a change reported as not statistically significant. A pre-registered validation harness contributed through the open process found that such shifts altered survey-answering behavior more than open-ended behavior. The milestone also produced version 0.1 requirements for the project’s data governance and management strategy, with work groups organized for both areas.

Milestone One and the 2026–2027 Roadmap

Tapestry contributors presented a paper, “Sovereignty Through Interdependence: Epistemic Agency, Intelligent Action, and Federated AI,” at the First Workshop on Sovereign AI for Collaborative and Pluralistic AI Ecosystems, co-located with ACM HCOMP 2026 in Washington, D.C.

“Our first milestone and growing global collaborations demonstrate that Tapestry is moving from architecture to working technology,” said Anthony Annunziata, co-founder and chair of the board of the AI Alliance.

Milestone One runs from September through November 2026 across consortium training, cultural alignment, and data governance and management, with OpenMined, Common Crawl, and the GSMA joining the work. The published roadmap lists Phase 1, a first version of the consortium training platform, as active now; a first base model is targeted for the end of 2026, early deployment for the first half of 2027, and a frontier-scale effort for summer 2027 or later. Movement from one phase to the next depends on technical proof points, partner commitments, compute availability, data readiness, and funding thresholds.