Milan/Brussels, January 16, 2025: The AI company iGenius announced today the launch of Colosseum 355B, a new state-of-the-art Large Language Model (LLM) with 355 billion parameters, designed to revolutionize AI capabilities in highly regulated industries. Built with the latest NVIDIA technology, Colosseum 355B enables regulated organizations to create regionally-specific LLMs with Continual Pre-Training (CPT). Additionally, it ​​makes model building at scale accessible to more enterprises, a capability that was previously limited to hyperscalers. 

The challenges of AI for regulated industries

Regulated industries have often faced complex AI challenges due to strict legislation and compliance requirements. Approximately 80% of the most valuable enterprise data cannot be exported and fine-tuned to centralized models and/or open models with a limited ownership license. Data such as personal information, financial transactions, trade secrets, and intellectual property cannot circulate outside of an organization’s network. As such, generative AI models can only be owned and controlled by governments and organizations that will use them with an isolated end-to-end experience. While centralized AI models/APIs are convenient for content generation and drafting, they are limiting for regulated industries. This is especially true of generative AI, which merges data and intellectual property irreversibly. For example, financial institutions sharing sensitive data with centralized LLMs can lead to potential data breaches or misuse, which can expose proprietary trading strategies, signal market intentions, and enable market manipulation.

A solution for regulated industries

One thing is clear, superintelligence must be decentralized to ensure the future of democracy and open markets. Colosseum 355B has helped push the boundaries of AI development for regulated industries.

Built using the NVIDIA AI Enterprise software platform, which includes NVIDIA NeMo, and leveraging the fully managed NVIDIA DGX Cloud AI platform, Colosseum 355B was developed on a single cluster with more than 3,000 NVIDIA H100 GPUs. This environment enables iGenius to rapidly scale its AI R&D and pre-training iterations, culminating in a highly capable model that supports over 50 languages, excels at coding, and is optimized for efficient deployment and resource utilization.

“Colosseum is a powerful AI model poised to unlock new opportunities for sovereign nations across the most highly regulated industries,”

said Alexis Bjorlin, vice president of DGX Cloud at NVIDIA. “Through collaboration with iGenius, NVIDIA AI experts helped optimize model training and provided access to NVIDIA AI and DGX Cloud fully managed computing clusters, enabling productivity from the start.” 

Colosseum 355B is designed for both CPT and fine-tuning. CPT allows enterprises to own proprietary LLMs by adding domain-specific knowledge and enabling long-term scalability, without sacrificing general knowledge. This way, organizations can own their AI brain, which will be existential for their future sovereignty. They can then use fine-tuning to create powerful task-specific adaptations. Colosseum 355B was pre-trained using FP8 precision to fit in one H100 GPU node, which makes CPT possible without requiring highly specialized AIOps skills. Pre-training this large-scale model in FP8 enables Colosseum to run on one H100 node, cutting inference cost by 50% without having to convert the model, which typically compromises accuracy or quality.     

Colosseum 355B is now available as an NVIDIA NIM microservice on the NVIDIA API catalog, offering organizations the opportunity to effortlessly integrate state-of-the-art AI capabilities into their operations. Future enhancements and use-case expansions are in the pipeline, paving the way for continued innovation and development.

For more information about Colosseum 355B, visit igenius.ai/language-models; to learn more about how Colosseum was built, read the NVIDIA technical blog here.