6GB of VRAM is quite limited for memory-intensive tasks such as vision-language models and makes it difficult to replace remote clusters.