Published January 25, 2026
| Version 1.0.0
Preprint
Open
SUVI: Scalable Unified Vector Intelligence for Efficient Edge Deployment
Description
Current Multimodal Large Language Models (MLLMs) are largely defined by ``alignment,'' the architectural compromise of grafting frozen, parameter-heavy visual encoders (e.g., CLIP, SigLIP) onto discrete language models. While effective, this dual-tower architecture imposes a prohibitive memory tax during inference, often requiring enterprise-grade GPUs to host both the language model and the high-dimensional visual encoder simultaneously. This creates a significant barrier to entry for deployment on resource-constrained edge devices and contradicts the growing need for sustainable computing. This paper introduces SUVI (Scalable Unified Vector Intelligence), a streamlined architecture that eliminates the need for a separate runtime visual encoder entirely. By rejecting continuous representation in favor of vector quantization, this approach forces all modalities into a shared, discrete embedding space, enabling a single Transformer to process multimodal data via a universal autoregressive objective. Empirical evaluations on the Flickr30k benchmark demonstrate that our unified framework retains competitive convergence behavior while reducing the inference memory footprint to 13.55 GB. This reduction allows the model to fit comfortably within the 16 GB envelope of commodity hardware (e.g., NVIDIA T4). By extending the lifespan of legacy infrastructure and reducing the operational energy required per inference, SUVI presents a viable pathway toward Green AI, democratizing high-fidelity multimodal intelligence while minimizing embodied carbon.