BIG DATA's KEYLM Intelligent Government Assistant Runs on NVIDIA's Integrated Hardware-Software Architecture

BIG DATA Co., Ltd., a leading data technology company in Taiwan, today announced the launch of KEYLM, an intelligent government assistant solution purpose-built for government decision support and legislative briefing. KEYLM brings together BIG10 (the company's proprietary domain language model trained on over 6.5 billion Traditional Chinese tokens accumulated over more than a decade) with an AI Agent platform, powered by NVIDIA DGX Spark as its on-premises compute foundation. By incorporating NVIDIA NeMo Framework and NVIDIA NeMo Agent Toolkit to strengthen model fine-tuning and Agent governance, KEYLM delivers a full-stack, single-device on-premises deployment that keeps responses traceable, processes auditable, and performance elevated.

"Sovereign AI demand is accelerating globally, and public sector agencies are under growing pressure to keep data on-premises while meeting increasingly strict security and compliance requirements," said Weiwei Chiang(蔣志薇), COO and Spokesperson of BIG DATA. "KEYLM runs on NVIDIA's integrated hardware-software architecture, and at the same time fills a longstanding gap: Traditional Chinese has historically been underrepresented in international AI training data. We built KEYLM to give government agencies an AI model that genuinely understands Taiwan's linguistic and institutional context, and solves three real operational problems: fragmented institutional knowledge, slow and inconsistent manual synthesis, and the inability to move sensitive data to the cloud."

According to Fortune Business Insights, the global sovereign cloud market is projected to grow from USD 195.35 billion in 2026 to USD 1.32 trillion by 2034, representing a compound annual growth rate of 27%. The firm identifies strong government and public sector demand for data residency and security governance as a primary growth driver. This trajectory makes one thing clear: on-premises deployment and data sovereignty have moved from niche government requirements to a mainstream challenge facing public institutions worldwide.

Taiwan's government agencies have long grappled with institutional knowledge scattered across official documents, meeting minutes, reports, and legislative briefing files. Retrieving the right information for an impromptu question or policy explanation often demands hours of manual searching. Manual synthesis is inconsistent and slow. Most critically, sensitive government data cannot leave the agency's controlled environment: it must remain on-premises, with responses that are fully traceable and processes that are fully auditable.

Addressing these challenges requires an AI model that genuinely understands Taiwan's local linguistic and institutional context. This is where the structural gap in global AI training data becomes a strategic issue. Take BLOOM, one of the most widely cited open-source multilingual large language models: English accounts for over 30% of its training corpus, and Simplified Chinese for approximately 16%, while Traditional Chinese is not independently tracked. The pattern is consistent across mainstream AI training data: Traditional Chinese has long been an afterthought.

BIG10, fine-tuned from a curated corpus of over 6.5 billion Traditional Chinese tokens built over more than a decade, was created precisely to close this gap. KEYLM's intelligent assistant capabilities go further. By integrating internal agency documents, operational data, and external public opinion signals, it delivers four core capabilities: instant Q&A, on-demand report generation, public opinion monitoring, and institutional knowledge management. Together, these transform the way government agencies access and act on information.

KEYLM is a fully integrated solution spanning three architecture layers: platform, model, and hardware. At the platform layer, KEYLM Agent Platform serves as the AI data hub for enterprise and government environments. It provides a four-tier system architecture, knowledge base management, traceable RAG-based Q&A, Agent workflow orchestration, long-term memory, and access governance with usage management. Agent governance runs on NVIDIA NeMo Agent Toolkit, enabling teams to build, run, and optimize AI Agent workflows with performance profiling, observability, and evaluation capabilities, keeping Agent behavior trackable, assessable, and auditable while reducing LLM call frequency, token consumption, and overall latency at the workflow level.

At the model layer, BIG10 is fine-tuned from over 6.5 billion curated Traditional Chinese tokens accumulated over more than a decade. It understands Taiwan's local linguistic conventions and government-specific terminology, and can switch flexibly between on-premises and cloud-based large language models depending on task sensitivity. Fine-tuning leverages NVIDIA NeMo Framework for data preparation, de-identification, domain adaptation, and evaluation, injecting Traditional Chinese domain knowledge directly into the model. De-identification capabilities are particularly critical for government use cases.

At the hardware layer, KEYLM Box ships with a single NVIDIA DGX Spark. With a NVIDIA GB10 Grace Blackwell Superchip and 128 GB unified memory allow one device to simultaneously run the AI Agent platform, the BIG10 model, and knowledge retrieval, making the entire intelligent government assistant stack fully self-contained on-premises. Following inference optimization on DGX Spark, KEYLM achieves approximately half the memory footprint and twice the response speed compared to unoptimized open-source models of equivalent scale.

Most enterprise AI deployments require substantial server infrastructure and dedicated data center space. KEYLM consolidates everything an AI-powered government assistant needs into a single desktop-sized DGX Spark. BIG DATA will continue building on its proprietary data assets and R&D capabilities, leveraging NVIDIA and other leading technology architectures to accelerate AI deployment for public sector agencies and enterprises. The goal is straightforward: AI applications that are contextually grounded, trustworthy, and built for Taiwan.