Self-hosting LingBot-Video on Apple Silicon: a 12-line diff off its CUDA-only stack, and what one clip actually costs

Context for why this belongs here: it is an open-weights, Apache-2.0 text-to-video model, so it is one of the few video generators you can actually self-host, instead of paying per-second on a hosted API like Runway, Pika, or Kling. The catch is that the official inference stack assumes an NVIDIA box top to bottom, so "self-hostable" on paper still meant "runs on an NVIDIA box I do not own." The 30B MoE version is out of reach on this machine. One base DiT in bf16 is about 60 GB, and a 64 GB Mac only expose