LM Studio hits SSD inference wall

This is less a bug report than a boundary test: LM Studio is great at making local LLMs approachable, but “run a model far bigger than memory from disk” remains a sharp-edge inference workflow.

• –LM Studio’s own docs describe model loading as allocating memory for weights and parameters, so expecting zero RAM pressure from the GUI is probably unrealistic.
• –The app uses llama.cpp and Apple MLX engines, but its UI does not necessarily expose every low-level llama.cpp flag or experimental loading pattern.
• –On Macs with unified memory, GPU/CPU layer sliders can blur what “RAM” means; disabling GPU offload does not make KV cache, buffers, metadata, or OS page cache disappear.
• –For giant MoE models like DeepSeek-style 671B-class GGUFs, community guidance still points to very large RAM budgets or direct llama.cpp experimentation, not LM Studio as a disk-streaming harness.