Provision of Canadian Large Language Model (LLM) for Inference

Objective of this project is to run additional Large Language Models locally in our existing infrastructure within the National Research Council Canada’s Azure Cloud. The Large Language Model should provide an API with a means to be queried against, and an appropriate response returned. There should be no data exfiltration, with all computation occurring locally. For data sovereignty and security, the provider of the LLM should primarily operate within Canada. The provided LLM is expected to be general use and meeting common industry benchmark metrics (MMLU, GPQA, IFEval, ChatRAGBench, StrategyQA, and others) within 5 percentage points or better of previous or current generation industry models (GPT-4o, Llama 3.3 70B, Mistral Large 2, DeepSeek V3, Gemini 2.0 Pro, etc.). The model should also be built and trained by the vendor from the ground up, without relying on tuning or re-training any existing models. The provider will be engaged for licensing and installation assistance. The NRC is exploring the deployment of Large Language Models (LLMs) within its secure cloud computing environment, seeking to operate LLMs locally and ensure that all data processing and computation occur entirely within the controlled environment, maintaining strict data sovereignty and security standards. Preliminary technical work has been completed to establish the foundation for this undertaking, including the setup of a GPU-based virtual machine that is able to communicate with the in-house developed LLM service delivery platform. This initiative aligns with NRC’s ongoing commitment to advancing secure, sovereign AI capabilities that support Canadian innovation, protect sensitive information, and ensure compliance with government data management standards.