LM Studio CPU Threads Peak at Five

This is a useful reminder that local LLM performance tuning is often bottlenecked by memory, not raw core count. Once you start mixing GPU offload with CPU-resident MoE layers, “more threads” can become counterproductive fast.

• –The post tests `qwen3.6-35b-a3b@MXFP4` with all GPU layers offloaded and 16 forced CPU layers, so the result is specific but practical
• –The drop-off above five threads lines up with the common RAM-bandwidth ceiling on consumer systems, especially older DDR4 platforms
• –The finding matters for LM Studio users because its MoE CPU-offload feature makes this tuning path easy to hit in real workloads
• –The discussion also reinforces that prompt processing and token generation behave differently, so one thread setting may not fit every phase
• –For developers serving local models, this is a reminder to benchmark thread pools per machine instead of assuming physical core count is the right target