
Local AI
Run a 120B MoE on an 8-24GB GPU with llama.cpp --n-cpu-moe
Moving MoE experts into system RAM can make a model five times faster, or noticeably slower. Both are true. With Qwen3.8 Flash Next now running on 8GB and 12GB cards, here is what the flag moves, where your RAM becomes the limit, and which speed tricks quietly change the answers.
Aliteq Team · · 12 min