Our Vision
Twinkle AI started with a simple conviction: language is the heartbeat of culture. If future digital tools can't grasp the nuances of how we actually speak, our identity risks being diluted by one-size-fits-all algorithms. Since early 2025, we’ve been curating local datasets and training models with one clear goal — to build an AI that truly 'gets' us and feels like a natural conversation partner.
How It All Started
Milestones
Founding Llama Taiwan
A small group of us gathered with a single mission: to build a model that speaks our language. This was the birth of the Llama Taiwan community.
Rebranded to Twinkle AI
We officially became Twinkle AI, symbolizing a community of bright minds coming together to build next-gen models.
Releasing Formosa-1
Collaborated with APMIC and National Center for High-Performance Computing to build Formosa-1, a 3B-parameter model built for the local context.
Cybersecurity in Action
CyCraft integrated Formosa-1 as a Guardrail for their XecGuard enterprise product.
Fineweb-edu-zhtw Project
Partnered with Taiwan Mobile to build educational datasets, sharpening the model's performance on local knowledge.
Releasing Formosa Vision Dataset
Partnered with the Taiwan Cultural Memory Bank to curate visual datasets, aligning the model's perception with local culture.
IT Matters Award
Winning the Open Source Contribution Award. This trophy belongs to the community — every single one of you. Enjoy!
1st Anniversary Meetup Party
Our first in-person meetup. Seeing all that energy in the room gave us the fuel to keep pushing forward.