Multimodal lightweighting

Nota, a Seoul-based AI optimization company, has applied its compression and optimization approach to multimodal AI models, not just large language models. The company reports the technology reduced GPU memory requirements by 47%, allowing multimodal models to run on lower-specification hardware.

Why it matters

Memory footprint is a key cost driver for deploying multimodal AI, which processes multiple data types such as images and text. A 47% memory saving could let companies run such models on consumer-grade GPUs or edge devices rather than high-end data center hardware.

Nota has previously focused on making AI models smaller and faster for real-world deployment, and the latest result suggests the same techniques now cover multimodal architectures. Further technical details of how the reduction was achieved were not provided in the source material.

The claim that Nota has brought lightweighting to multimodal AI with a 47% cut in GPU memory use is mostly true.