The author details the development of Marvis AI Studio, a localized AI video generation service hosted on a single consumer-grade NVIDIA RTX 5060 Ti 16GB GPU. To bypass cloud computing costs, the developer implemented a full stack consisting of ComfyUI as the generation engine, the MiniMax H3 video model using pruned INT8 quantization to fit within the 16GB VRAM limit, and the qwen3-VL (32B) vision-language model utilizing NVFP4/AWQ for prompt understanding. Public access is managed via a Python-based API gateway running on port 8088, exposed through a Cloudflare tunnel to avoid port forwarding. Performance benchmarks reveal significant latency inherent to consumer hardware: generating a 3-second video at 864×480 takes approximately 6 minutes, while a 5-second video at 1664×928 requires roughly 40 minutes. To ensure stability, the author resolved ComfyUI execution errors by using the --cuda-device 0 and --disable-pinned-memory flags and extended gateway timeouts to 60 minutes to accommodate long processing cycles. The service employs a tiered access model, offering three free generations per IP per day to mitigate bot abuse, while paid access is managed via redemption codes. Ultimately, the project demonstrates the feasibility of running low-scale, hobbyist AI services by prioritizing specific quantization methods to match hardware VRAM constraints.