IQ4 qwen 122b-a10b would mean 61GB total size and 5GB active, so about 5GB of the model loaded into GPURAM plus any generated context, and 61GB of weights loade

IQ4 qwen 122b-a10b would mean 61GB total size and 5GB active, so about 5GB of the model loaded into GPURAM plus any generated context, and 61GB of weights loaded into system RAM? I don't know if that math is correct, but does that run well? Wouldn't that only leave 3GB of system RAM?