AirLLM loads one layer at a time, so giant open models run on small GPUs (Apache-2.0). Its README lists Kimi K3 (2.8T) in 3.72 GB. The catch: it's slow.