Research
At APMIC, we are committed to advancing AI technology through research and innovation, with a strong focus on private Large Language Models (LLMs) and Fine-tuning. Our mission goes beyond developing customized AI solutions for businesses; we strive to make our research accessible to the public, simplifying complex AI concepts into powerful, bite-sized tools that anyone can leverage effectively.
With a decade of deep expertise in AI language models
Since our founding in 2017, we have remained focused on one core mission: building vertical sovereign AI Models tailored for industry needs.
Key Milestones:
-
2017–2022: APMIC was founded and became an early mover in the field of Natural Language Processing, conducting in-depth research on world-class technologies such as BERT, GPT-2, and PaLM.
-
2023–2024: The CaiGunn model was launched, introducing a Mixture of Experts (MoE) architecture that delivered a significant leap in performance.
-
2025 and beyond: The ACE Traditional Chinese model series was released, redefining enterprise AI applications through lighter-weight architectures, higher precision, and reasoning capabilities designed specifically for real-world enterprise use.
CaiGunn 34B
Traditional Chinese Language Model for All Use Cases
In January 2024, CaiGunn 34B ranked 64th globally on the Open LLM Leaderboard hosted by Hugging Face, with an average score of 71.19. In Taiwan, CaiGunn 34B also claimed the top position, marking a significant milestone for the Traditional Chinese language model in the APAC region.
CaiGunn 34B is built on APMIC's Brainformers architecture, integrating the LLaMA model foundation with Mamba dynamic computation flows, the Transformer framework, and the Mixture of Decoding Experts (MoDE) technology. This combination delivers high precision and flexible generative capabilities. Running entirely on the NVIDIA NeMo Framework, the model features efficient distributed training and inference deployment, supporting enterprises' needs for on-premises integration and private applications.
ACE-1 Series
APMIC's Traditional Chinese Inference and Multimodal Models
ACE-1-3B is Taiwan’s first 3B-parameter Traditional Chinese language model capable of running directly on mobile devices. Through support for the Model Context Protocol (MCP), the model can connect to both internal and external enterprise systems, enabling natural language–driven information retrieval, workflow execution, and decision support. This creates a new generation of intelligent human–machine interaction interfaces. ACE-1-3B also supports integration with external APIs, ERP systems, and IoT devices.
Combined with APMIC’s PrivModel fine-tuning, distillation, and model foundry services, enterprises can easily build proprietary on-premise models tailored to their specific needs. These models can be smoothly deployed on mid- to low-tier GPUs or embedded platforms while maintaining real-time responsiveness. This significantly lowers both the technical and cost barriers to AI adoption, meeting the demands of industrial equipment manufacturers, technology companies, mobile device brands, and robotics and human–machine interface (HMI) applications that require edge deployment, low power consumption, and fast response times.
Media Coverage
.png)
【模型訓練技術論文】使用 NVIDIA NeMo 進行多節點指令微調：擴展 Qwen2.5-32B 並重塑 s1 模型
APMIC 研發部門研究員 / LLM 工程師 Ethan Kuo
APMIC 內部以 NVIDIA NeMo 進行跨節點分散式訓練，沿用 s1K 的 1,000 筆數理推理資料，對 Qwen2.5-32B-Instruct 進行微調以重現論文觀察。由於當時（2025/03）NeMo 正由 1.0 遷移至 2.0，且 NeMo 2.0 尚未完整提供 Qwen2.5-32B 的現成 recipe，我們手動修補了訓練配置與流程，並在 16 張 H100 的跨節點環境完成訓練。本篇文章將整理環境與設定修補重點、訓練流程，以及微調前後在 benchmark 上的結果與觀察。
Fine-tuning the FunctionGemma Model Using a DevOps Workflow and Deploying It as Google ADK Agents
APMIC MLOps Engineer — Simon Liu
Simon Liu was invited to speak at the Twinkle AI 1st Anniversary Offline Event and GDG Build with AI 2026, where he shared an in-depth introduction to the FunctionGemma model, its implementation workflow, demonstrated results, and practical conclusions. The session focused on real-world experiences in bringing AI development from experimentation to production deployment.
[DevOps in AI Agents] Evaluating Production Readiness for AI Agents —Google ADK AI Agent Evaluation: Technical Implementation (Part 2/2)
APMIC MLOps Engineer — Simon Liu
In this article, I will walk readers through the implementation of ADK Evaluation from three key perspectives to assess whether an AI Agent is ready for production deployment. For aspects not covered in this article, or in cases where updates have been released by the official team, readers are encouraged to refer to the official ADK Evaluation documentation for the most up-to-date information.
[DevOps in AI Agents] Evaluating Production Readiness for AI Agents —
Google ADK AI Agent Evaluation: Conceptual Overview (Part 1/2)
APMIC MLOps Engineer — Simon Liu
This article provides an in-depth introduction to the design philosophy and core functionalities of Google ADK Evaluation, explaining how it helps developers maintain quality control and development confidence for AI Agents, even in environments characterized by probabilistic behavior and variability.
Welcome to the Gemmaverse: Exploring the History of Gemma and a Comparison of the Latest Gemma 3n Models Unveiled at Google I/O
APMIC Founder & CEO — Jerry Wu
At the I/O Extended Taipei event, APMIC Founder and CEO Jerry Wu delivered an in-depth overview of the evolution of the Gemma model family. He also analyzed the newly released Gemma 3n alongside related models, including TxGemma, SignGemma, and DolphinGemma, highlighting their application differences and potential across various real-world scenarios following their announcement at Google I/O.
Rapidly build multimodal application prototypes using Google AI technology
APMIC MLOps Engineer — Simon Liu
Simon Liu, an APMIC MLOps engineer, was invited to participate in Google Cloud Summit Taipei, where he demonstrated how to leverage the powerful Google Gemini AI model to quickly create multimodal application prototypes using natural language. Whether it's images, text, or structured data, a single dialogue can transform concepts, making innovations more efficient to implement.
.jpg)
Knowledge Distillation in Enterprise AI
APMIC Founder & CEO — Jerry Wu
APMIC Founder and CEO Jerry Wu was invited to speak at the 2025 Generative AI Developer Conference, where he shared how model fine-tuning and knowledge distillation techniques address real-world enterprise deployment challenges and enable organizations to build truly sovereign AI systems. The session received strong engagement and active discussion from the developer community.
Through knowledge distillation and testing phases, LLM accuracy and computational efficiency are improved.
APMIC Founder & CEO — Jerry Wu
APMIC CEO Jerry was invited to speak at GTC Taipei, where he shared our firsthand experience in AI model refinement and commercial applications, which generated a warm response from the audience and facilitated in-depth dialogue with the international technology community, exploring more possibilities for cooperation.
How Should Enterprises Deploy Private LLMs? APMIC and Advantech Reveal the Key to Real-World AI Deployment
Eli, co-founder and head of product at APMIC
Invited by Advantech to participate in the Digitimes seminar, APMIC shared their firsthand experience in implementing privatized LLM in manufacturing, finance, and government agencies, providing in-depth analysis on: Why is "knowledge management" the most suitable first step for LLM? How does APMIC enhance response quality using self-trained models, data distillation, and fine-tuning techniques?