Most people assume the biggest models need the most VRAM! Actually, with AirLLM, Kimi K3, a 2.8 trillion parameter model, needs less VRAM than a 70B dense model. The reason is that K3 is a sparse mixture-of-experts model, so AirLLM streams only the specific experts a token actually routes to, instead of loading a full dense layer. Combined with loading just one layer onto the GPU at a time. That's how DeepSeek-V3 (671B) fits in 12GB and a 2.8 trillion parameter model fits in under 4GB, with no quantization, no distillation, and no pruning involved. Bookmark so you do not lose it! Repo: https://t.co/dO2TABPfqr Follow @neil_xbt for more!