Case Study: Unum preserves knowledge in compact AI models with Nebius

How Unum partnered with Nebius to preserve knowledge in compact models

Unum, an AI research lab specializing in compact multimodal models, faced significant challenges in training due to the immense data loading speeds required and the difficulty of preserving knowledge within their small, efficient architectures. The fluctuating metrics and high I/O-to-computing power ratios ruled out virtualized storage, creating a bottleneck. They partnered with vendor Nebius to leverage its Compute Cloud with high-performance NVIDIA H100 GPUs to address these constraints.

Nebius provided the scalable, high-performance computing infrastructure necessary for Unum's intensive training processes. This partnership enabled Unum to successfully train and open-source multiple advanced models, including UForm-Gen2. The results demonstrated superior efficiency; for instance, Unum's models trained on a smaller but higher-quality dataset of 25 million samples outperformed OpenAI's CLIP model, which was trained on 400 million pairs. The collaboration with Nebius provided the computational power needed to advance compact, data-efficient AI that can run on devices like mobile phones while preserving user privacy.

Unum