You need to agree to share your contact information to access this model
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Access to this model requires manual review by APMIC. Please provide your company name and contact email so we can process your request. 本模型採人工審核制，請填寫公司名稱與聯絡 Email 以利審核。
Log in or Sign Up to review the conditions and access this model content.
ACE-1-24B-NVFP4
Model Description
ACE-1-24B-NVFP4 is the NVFP4-precision release of ACE-1-24B, APMIC's flagship self-developed Traditional Chinese reasoning model, built for enterprise applications with a 65K long-context window. ACE-1-24B is trained on three major Traditional Chinese data pillars — mathematical & logical reasoning, everyday commonsense reasoning, and tool-use instructions — and natively supports Chain-of-Thought (CoT) reasoning chains, enabling stable, verifiable responses and knowledge integration in on-premise environments.
This release demonstrates APMIC's end-to-end capability in reasoning model development, localization, and hardware–software co-optimization using NVIDIA precision formats, delivering production-ready AI aligned with modern GPU infrastructure.
Model Details
- Developed by: Min Yi Chen、Liang Hsun Huang、Wen Bin Lin & Dave Sung (All authors have contributed equally to this work.)
- Funded by: APMIC, under the leadership of CEO Jerry Wu
- Model type: Causal decoder-only Transformer, 24B parameters (reasoning model)
- Context length: 65K tokens
- Language(s) (NLP): Traditional Chinese & English
- License: APMIC proprietary license (gated on Hugging Face; access granted via manual review)
Reasoning Capabilities
Chain-of-Thought Native Reasoning
ACE-1-24B is trained as a reasoning-first model: it decomposes problems into explicit intermediate reasoning steps before producing a final answer, improving reliability on multi-step tasks. Its training corpus is organized around three Traditional Chinese data pillars:
- Mathematical & logical reasoning — arithmetic, symbolic, and structured logical problem solving
- Everyday commonsense reasoning — real-world inference grounded in Taiwan-centric linguistic and cultural context
- Tool-use instructions — function calling and instruction patterns for agentic and workflow automation scenarios
This design yields:
- Transparent, auditable reasoning chains for enterprise review
- Robust multi-step problem solving in native Traditional Chinese
- Reliable structured outputs for tool invocation and system integration
65K Long-Context Understanding
With a 65K-token context window, ACE-1-24B sustains coherent reasoning across long documents and multi-turn sessions, enabling:
- Long-form enterprise document comprehension and cross-referencing
- Retrieval-augmented and knowledge-integration workflows
- Extended agentic sessions with persistent task state
NVIDIA NVFP4 Precision Optimization
NVFP4 Quantization for Next-Generation Inference
This release converts ACE-1-24B to NVFP4 precision, leveraging NVIDIA's hardware-native numerical format and software toolchain. Through tight integration with NVIDIA's inference ecosystem, APMIC achieves:
- Major reductions in memory footprint and bandwidth usage
- Significant gains in inference throughput and energy efficiency
- Preservation of reasoning quality and instruction performance
- Production readiness for large-scale enterprise AI services
This highlights APMIC's capability in NVIDIA-aligned precision engineering and deployment optimization.
Hardware and Deployment Efficiency
Built for On-Premise NVIDIA AI Infrastructure
With NVFP4 precision and deployment-aware optimization, APMIC/ACE-1-24B-NVFP4 delivers:
- Ultra-efficient inference on modern NVIDIA GPU architectures
- Compatibility with NVIDIA runtime and acceleration libraries
- Stable, secure deployment in on-premise and private cloud environments
- Reduced total cost of ownership for enterprise reasoning workloads
Positioning
This model represents APMIC's capability to deliver self-developed, NVIDIA-optimized, enterprise-grade reasoning AI systems through a complete lifecycle of:
reasoning-focused training → Traditional Chinese localization → NVIDIA precision optimization.
It is intended for organizations requiring:
- Advanced Traditional Chinese reasoning and Chain-of-Thought intelligence
- 65K long-context comprehension for document-heavy workflows
- Maximum efficiency on NVIDIA GPU infrastructure
- Secure, scalable, on-premise AI deployment with stable responses and knowledge integration