Our client, the Finnish education startup specializes in digital learning platforms and intelligent educational content. With the rapid development of AI technology, the company plans to deploy a local DeepSeek platform to support the research and development for its next-generation intelligent teaching and personalized learning systems.

Originally focused on educational content, the company lacks experience in deploying high-performance computing (HPC) and large language models (LLMs). They faced significant challenges, particularly in selecting the appropriate DeepSeek model version, planning hardware resources, and designing the network architecture. Our client aims to build a sustainable, scalable enterprise-grade AI computing platform in a single implementation, laying the foundation for future AI-driven educational applications.

The client's team had no prior experience with large language models like DeepSeek, lacking familiarity with model training, inference architecture, and computing power requirements. They also had limited systematic technical selection capabilities.

While the education company had application-layer development experience, it lacked expertise in server hardware, storage architecture, network deployment, and computing optimization—requiring a vendor-provided, fully integrated solution.

The client maintained long-term partnerships with local network equipment suppliers and mandated the exclusive use of Arista-branded networking devices, imposing higher demands on solution compatibility and design flexibility.

After thoroughly analyzing the client's educational use cases and DeepSeek platform requirements, AICPLIGHT designed and deployed a high-performance, low-latency, cost-optimized enterprise private network solution to support the on-premises deployment of the DeepSeek large language model.

Through in-depth evaluation of the client's target applications—including automated teaching assistants, multi-disciplinary question banks, and AI-powered scheduling, AICPLIGHT determined the computing demand at approximately 800B parameters. Consequently, AICPLIGHT deployed 12 servers with DeepSeek-R1-70B models to balance parallel computing and inference capabilities.

This architecture achieves high-bandwidth isolation and independent resource scheduling between compute and storage networks. It maintains efficient and stable system operation while providing excellent scalability for future node expansion.

The entire network solution primarily uses AOC (Active Optical Cables) as the transmission medium, reducing overall costs by approximately 25% compared to traditional optical transceivers with fiber patch cords.

AICPLIGHT offers customized AOC cables (1m to 100m) to flexibly adapt to rack layouts, minimizing cable redundancy and complexity. Standard lengths (1m/5m/10m/20m/30m) are stocked for rapid delivery.

Tailored to the client's education needs, the DeepSeek-R1-70B model balances performance and resource efficiency—meeting the 800B-parameter compute target while reserving capacity for future fine-tuning and local knowledge training.

The solution achieves comprehensive integration across hardware, networking, and model layers. The deployment process is transparent, eliminating the need for customers to assemble complex components themselves. Upon launch, the system supports AI functionalities including teaching Q&A, courseware generation, and learning path recommendations, significantly boosting educational content production efficiency.

The solution reserves expansion ports and compute node slots, enabling seamless future scaling through GPU node additions or storage bandwidth upgrades. AICPLIGHT provides full lifecycle technical support covering model tuning, computing power monitoring, and network maintenance.