An MCP Service Bridging Local Data and AI Workflows

Twinkle Hub is an MCP (Model Context Protocol) service designed specifically to connect Taiwan's open government data with modern AI workflows. Through a standardized interface, it transforms complex and diverse government datasets into context that AI models can directly read and process.

Moving Beyond Manual Data Ingestion

Traditionally, developers have had to manually import data into AI systems, a process that is inefficient and prone to outdated information. Twinkle Hub was created to solve this bottleneck. With its precise data retrieval mechanism, AI models can accurately search and extract content from public government databases—including government procurement records, national examination questions, or real-time housing price data—significantly improving data utilization.

Automated Workflows with Daily Updates

More importantly, the system employs a daily update mechanism, ensuring that AI models always access the latest open data when executing tasks and eliminating the risks associated with static, obsolete information. At the application level, developers can leverage Twinkle Hub's dynamic data integration to enable AI agents to perform automated analysis and report generation, effectively replacing the cumbersome manual procedures of the past.

A Robust Foundation for Data Integration

Twinkle Hub currently integrates 49,000 government open datasets. This technical architecture not only strengthens AI's understanding of regional information but also serves as vital support for local application development, providing the technical community in Taiwan with a stable and contextual data source.

Government Support for AI Through Open Data: Recognizing Excellence in Text Corpus Development

The Ministry of Digital Affairs (moda) has officially presented its first "Outstanding Open Text Corpus Awards," marking a significant milestone in the government's open data policy to support artificial intelligence development. This initiative recognizes organizations that have demonstrated excellence in the public release and quality maintenance of text data, providing reliable and representative training materials for the developer community.

The establishment of this award reflects the government's commitment to the intersection of open data and AI development. By encouraging organizations to proactively release high-quality, legally licensed Traditional Chinese text datasets, this initiative helps improve the training efficiency and logical reasoning capabilities of local AI models. For development communities like Twinkle AI, which focus on model training and open-source technology, this represents a crucial resource for enhancing overall technical competitiveness.

Turning Passion into Protocol

January 2026 marked a defining moment for Twinkle AI. This wasn’t just an anniversary; it was our first official physical meetup — the day we took our technical syncs out of Discord and turned our online energy into real-world impact.

Our Evolution: From Feedback to Framework

We originally hit the scene as Llama Taiwan on January 2, 2025, with a simple mission: gather feedback for the Llama-3.2-Taiwan model. But as the community grew, so did our ambitions. We quickly evolved into a dedicated technical hub for Traditional Chinese AI models, bringing together the brightest minds to build out the local open-source ecosystem.

70+ Strong: A Community Without Borders

We initially prepped for 40 friends, but 70+ passionate partners showed up. Seeing members travel all the way from Kaohsiung to Taipei just to talk tech — that’s the kind of community spirit that keeps us going.

Event Lowdown:

• The Roadmap Ahead: We looked back at a year of collective building and laid out the blueprint for where we’re heading next.
• Beyond the Screen: We finally turned those Discord avatars into real-life collaborators, building the kind of deep-rooted partnerships that can't be forged through a screen.
• Lightning Talks: From deep-dives into model training to the latest in AI Agent implementation, our contributors showcased the raw technical expertise that defines this community.

A massive thank you to everyone who’s been part of this journey. We’ve faced our fair share of challenges, but the people who stayed are here because they believe in the power of open source. Every breakthrough we've achieved is the result of collective grind — a foundation built by the hands of every single contributor.

"This Win belongs to Every One of You"

Twinkle AI just hit a major milestone at the IT Matters Awards, hosted by the IMA (Information Manager Association). We’re incredibly proud to take home the Open Source Community Contribution Award. This achievement was driven by our tight collaboration with Taiwan Mobile on open-source initiatives, where we joined forces to turn months of hard work into a winning submission.

Sharing the stage with veteran communities like PyCon and SITCON was a massive reality check and a huge confidence boost. This award isn’t just a trophy; it’s a validation of our technical grind and a bold step toward our goal of building a resilient, local AI ecosystem.

"This trophy belongs to every single partner. Your energy is why this community was recognized. The glory is all yours. Enjoy it!"

Connecting with so many OG communities reminded us that Twinkle AI is still just getting started. This win gives us the momentum to tackle the real challenges ahead — figuring out talent retention, streamlining knowledge transfer, and building the kind of deep-rooted partnerships that keep our technical energy alive.

The Twinkle AI team officially unveiled Formosa Vision, a project designed to address the cultural blind spots of global Vision Language Models (VLMs). We curated thousands of image-text pairs from the National Cultural Memory Bank, applying precise Data Cleaning and linguistic calibration to ensure AI can accurately recognize unique Taiwanese landscapes—from historic trails and military housing to the battlefields of Matsu.

We showcased our technical results at the TAIWAN AI RAP event organized by the NSTC and NARLabs. In collaboration with Chen-Si Studio, we demonstrated Taiwan’s first small-scale reasoning model, proving that edge-device AI has massive potential for everyday logic and tasks.

Topic: GPT-OSS: A Deep Dive into Hardware Requirements from Training to Inference

At the September Generative AI community meetup, we shared a practical developer’s perspective on the hardware architecture needed for open-source models. The session covered everything from the initial training phase to backend inference deployment, providing developers with a practical roadmap for choosing computing hardware and planning their setup.

• Time: 2025/09/17 (Wed) 19:40 - 19:50
• Location: SPARKFUL (12F, No. 1-1, Sec. 1, Chongqing N. Rd., Datong Dist., Taipei City)

Topic: Art of GRPO Training (Lightning talk)

At Taiwan's premier annual AI gathering at Academia Sinica, we took the stage for a lightning talk on the Art of GRPO Training. We focused on how Reinforcement Learning can be used to sharpen a model’s logical reasoning, while swapping technical notes on the fine-tuning process with experts and fellow researchers across the field.

• Time: 2025/09/10 (三) 15:16 - 15:28
• Location: International Conference Hall, HSSB, Academia Sinica

Under the guidance of the Ministry of Digital Affairs (MODA) for the "AI Industry Ecosystem and Multi-Talent Empowerment Program," Twinkle AI was invited to lead a technical session. We joined partners across industry, government, and academia to share our fine-tuning insights and practical development experiences, exploring ways to build a more resilient and complete local AI technical ecosystem.

Time and Location

• Event Date: 2025/07/30 (Wed)
• Event Time: 13:30 ~ 16:00
• Location: No. 287, Sec. 3, Chengde Rd., Datong Dist., Taipei City

Program Framework

• Advisory: Ministry of Digital Affairs (MODA)
• Organizer: Administration for Digital Industries (ADI)
• Co-organizers:
  • Institute for Information Industry (III), ALC Taipei (Apache Software Foundation)
  • National University of Kaohsiung (NUK), National Taitung University (NTTU)
  • Institute of Information Science, Academia Sinica
  • Twinkle AI Community
  • National Pingtung University of Science and Technology (NPUST)
  • Open Culture Foundation (OCF)
  • PostgreSQL Taiwan User Group

Agentic AI x Blue Team x Red Team

The 3B-parameter open-source model F1, co-developed by Twinkle AI and APMIC, was invited to serve as the testing foundation for XecGuard, the latest AI security technology from CyCraft. Through live-fire stress tests, we demonstrated the exceptional resilience of local models when paired with security fine-tuning modules against malicious attacks.

• Time: 2025/07/04 09:30 - 17:00
• Location: Great Skyview, 13F, Glair Hall (No. 101, Bangka Blvd., Wanhua Dist., Taipei City)

CyCraft and APMIC entered a strategic partnership to unveil XecGuard, the industry's first defense-oriented AI security module. Powered by a proprietary multi-stage fine-tuning process (SFT+GRPO), this "plug-and-play" solution provides essential security for AI systems, helping organizations defend against malicious attacks during localized deployments.

In technical trials, the R&D team used F1 (Formosa-1)—the 3B model co-developed with Twinkle AI—as the testing benchmark. Verified through LLM Red Teaming, the results demonstrated exceptional defensive capabilities:

• Measurable Security Boost: Overall security scores increased by an average of 19.4%, with specific attack scenarios showing up to a 33.9% improvement.
• Outperforming Global Models: The F1 + XecGuard combination significantly outperformed international commercial models, including Llama 3B, Gemma 4B, and DeepSeek 8B, in security resilience.
• Precision Defense: Effectively counters threats like Prompt Injection and Prompt Extraction while maintaining high-quality responses for essential tasks.

In collaboration with APMIC, we introduced Formosa-1, Taiwan’s first 3B reasoning model capable of running on mobile devices. Leveraging Model Distillation and Taiwanese Chain-of-Thought (TCoT) data, we significantly enhanced the model’s performance in legal reasoning and logical deduction.

Alongside the model, we launched Twinkle Eval, an open-source evaluation framework. By implementing massive parallel testing and randomization, it ensures stability across local benchmarks like TMMLU+ and specialized legal datasets, providing a fairer and more precise standard for Traditional Chinese AI development.