APMIC releases the ACE series open-source models 4B and 12B: combining NVIDIA NVFP4 and knowledge evaporation technology, defining a new benchmark for enterprise-grade AI assistants.

March 9, 2026 at 1:00:00 PM

APMIC has officially released four ACE-gemma-3 models, fine-tuned for the Taiwanese context and combining knowledge distillation and NVFP4 inference technology, providing enterprises with lightweight 4B and 12B model specifications that balance privacy, autonomy, and efficiency.

APMIC, a leading brand in enterprise sovereign AI solutions, today officially announced the open-sourcing of its four deeply optimized ACE-gemma-3 series large language models. Developed by APMIC's elite AI modeling team using the PrivModel fine-tuning and distillation solution, this series is not only deeply fine-tuned for Taiwan's local linguistic context but also leverages knowledge distillation technology alongside NVIDIA's next-generation NVFP4 inference precision. Available in 4B and 12B lightweight yet highly efficient specifications, these models are designed to help enterprises construct proprietary AI assistants while ensuring data privacy and technological sovereignty.

Core Technology: Strategic Division Between 4B & 12B to Meet Diverse Commercial Scenarios

"True 'Enterprise Sovereign AI' should not be constrained by exorbitant computing costs," said Jerry Wu (Po-Han Wu), Founder and CEO of APMIC. "During the model training process, we specifically optimized the weights for both the 4B and 12B versions, aiming to allow enterprises to flexibly select the most suitable 'AI brain' based on the complexity of their business tasks."

ACE-12B: Built for High Logic, Complex Instruction-Following, and Cross-Lingual Tasks Designed for tasks requiring high logical reasoning, complex instruction-following, and cross-lingual processing. Through its 12B (12 billion parameters) depth combined with NVFP4 quantization technology, ACE-12B can process enterprise-grade long texts, automate workflows, and provide precise decision-making assistance at an extremely high throughput. It is the ideal choice for building internal private Retrieval-Augmented Generation (RAG) knowledge bases and high-end digital assistants.

ACE-4B: Exceptional Efficiency Driven by Advanced Knowledge Distillation

The 4B specification showcases APMIC's cutting-edge knowledge distillation technology. Despite its lightweight size, its performance on specific tasks rivals that of larger models, boasting astonishing inference speeds. The 4B (4 billion parameters) model is highly suited for deployment in real-time customer service systems, mobile applications, or edge computing devices, significantly reducing the Total Cost of Ownership (TCO) for concurrent processing and enabling scalable AI applications.

According to the latest internal MMLU and TMMLU+ mathematical benchmark results, ACE-4B exhibits extraordinary accuracy in the mathematical domain, even surpassing Gemma-27B, a model over six times its parameter size. Across various subjects including "Elementary Mathematics," "Junior High School Mathematics," "Senior High School Mathematics," "Vocational Mathematics," "Engineering Mathematics," and "Abstract Algebra," ACE-4B demonstrates exceptionally robust logical reasoning capabilities.

Leading the Industry! The Perfect Synergy of NVIDIA NVFP4 with Localized Pre-training and Fine-tuning

The ACE series models not only possess robust hardware-aware capabilities but also demonstrate deep Taiwanese localization in semantic understanding:

NVIDIA NVFP4 Ultimate Acceleration: APMIC takes the lead in supporting NVIDIA’s next-generation low-bit numerical format. Through NVFP4 optimization, the model's memory footprint on NVIDIA GPUs is significantly reduced. This enables enterprises to achieve faster response speeds with fewer resources, directly transforming computing efficiency into tangible commercial competitiveness.

Pre-training and Fine-tuning Tailored for Taiwan's Linguistic Context: The training process incorporated a vast corpus of local Taiwanese legal, financial, and business data. This ensures that when the AI assistant generates responses, it aligns perfectly with Taiwan's unique tone and cultural norms, achieving true localized AI sovereignty.

Model	Key Highlights	Commercial Value
ACE-gemma-3-12b-it-nvfp4	12B Parameters / NVFP4 Optimization	High-complexity enterprise-grade AI assistants that support long-text analysis, automated workflows, and cross-lingual decision assistance.
ACE-gemma-3-12b-it-fp8	12B Parameters / FP8 Precision	Optimized for modern data centers to deliver ultra-stable generation quality, making it ideal for highly regulated industries such as finance and healthcare.
ACE-gemma-3-4b-it-nvfp4	4B Parameters / NVFP4 Optimization	The premier choice for lightweight deployment. Retains powerful reasoning capabilities through knowledge distillation, making it ideal for real-time customer service and edge computing devices.
ACE-gemma-3-4b-it-fp8	4B Parameters / Localized Fine-tuning	Perfect-fit alignment with Taiwan's cultural context, making it ideal for brand interaction tools that require high approachability and precise semantics.

Experience It Now: The ACE Series Models Are Officially Live on Hugging Face

To accelerate the adoption of enterprise sovereign AI, APMIC has made all of the aforementioned models available on Hugging Face, the world’s leading AI community platform. Enterprise developers and researchers can download the model weights immediately to explore their exceptional quantization performance and localized inference capabilities.

Download and learn more today: https://huggingface.co/APMIC

Looking for the AI Model That Best Fits Your Business Needs? Choose PrivModel, the Tailored Enterprise Model Solution.

To meet the precise demands of enterprises for domain-specific knowledge, we have launched the professional PrivModel solution .Through this service, we can further fine-tune models using our enterprise clients' proprietary, specialized internal data. Concurrently, by leveraging distillation technology, we help reduce model size and optimize deployment efficiency. This ensures that the model transforms into a proprietary, domain-specific asset that yields the best possible performance internally, ultimately aligning model capabilities perfectly with practical business value.

Be the first to try the PrivModel Solution