GB300 Coming soon. Learn more

December, 2025 | Partners

AI Singapore x Firmus - Case Study

Advancing regional AI innovation with high-performance, efficient infrastructure.

AI Singapore (AISG), a national AI programme launched by Singapore’s National Research Foundation, partnered with Firmus Technologies to support the development of SEA-LION, the first family of open-source large language models designed specifically for Southeast Asian languages and contexts.

Through Firmus’ AI cloud platform, AISG gained access to reliable, high-performance compute infrastructure that enabled rapid experimentation, large-scale training, and efficient model evaluation, accelerating cutting-edge AI research tailored to Southeast Asia’s unique linguistic and cultural diversity.

The challenge

Prior to working with Firmus, AISG faced major challenges in securing guaranteed, large-scale, energy-efficient GPU clusters for advanced LLM training. These challenges were compounded by high operational costs and infrastructure complexity associated with Singapore’s demanding climate.

Scaling experimentation while maintaining performance, efficiency, and reliability posed significant constraints on model development timelines and research throughput.

The solution

We addressed these challenges by providing a purpose-built, AI-first infrastructure. Designed specifically for demanding AI workloads, the platform delivered consistent performance and operational reliability, supported by a highly responsive and proactive engineering team.

This allowed AISG researchers to accelerate large-scale experimentation and model development without compromise, even under intensive training and evaluation cycles.

Key outcomes and performance

SEA-LION enables developers to build accurate, regionally relevant generative AI solutions, from multilingual customer support to advanced language analytics, fostering broader AI adoption across Southeast Asia.

Key statistics achieved during the engagement include:

  • 32 nodes / 256 NVIDIA H200 GPUs deployed
  • NCCL bandwidth of 317 GB/s and throughput of 387 TFLOP per device per second
  • Over 200 experiments completed
  • A 27B parameter model trained in 10 days using 32 nodes
  • A 4B parameter model trained in 3.5 days using 16 nodes
  • More than 100 candidate models produced and evaluated
  • Supervised fine-tuning runs completed in 4 hours
  • Each checkpoint evaluated in 1.5 hours using 2 nodes
AI Singapore x Firmus key statistics of GPU AI Cloud infrastructure

Strategic partnership

AISG and Firmus formalised their collaboration through a Memorandum of Understanding, establishing a long-term partnership to redefine AI research capabilities and accelerate innovation across Southeast Asia.

The partnership focuses on three core initiatives:

  1. Access to Singapore-based H200 GPUs hosted in Firmus AI Cloud’s Sustainable AI Factory, leveraging next-generation immersion cooling to significantly reduce operational costs and energy consumption.
  2. Global hosting of SEA-LION, ensuring scalable, accessible, and environmentally conscious deployment of regionally focused AI models.
  3. Joint AI benchmarking research, evaluating AI training and inference workloads across compute configurations to optimise performance and unlock deeper industry insights.

Impact

By combining AISG’s mission to anchor deep national AI capabilities with our sustainable, high-performance infrastructure, the collaboration supports the creation of new knowledge, regional AI leadership, and environmentally responsible innovation.

The partnership reflects a shared commitment to advancing AI adoption across Southeast Asia while enabling researchers and developers to move faster, from research to production, on a global stage.