Developer Runs 120B MoE Models on M1 Max With 16 GiB Memory Budget
A developer kept MoE expert weights on SSD and loaded them on demand, running gpt-oss-120b and Qwen3.5-122B-A10B within a 16 GiB process budget on a 32GB M1 Max. On gpt-oss-120b it averaged 7.84 tokens/s, about 36% faster than TUFF, and degraded less under competing memory load.