How to Calculate Actual RAM Needs for Running LLMs Locally on a Mac

Running large language models locally on a Mac requires accounting for two distinct memory components: fixed weight memory and a growing KV cache. Weight memory is determined by parameter count multiplied by bytes-per-parameter, meaning a 7B model at 4-bit quantization uses roughly 3.5GB for weights alone. The KV cache, however, scales with conversation length and can rival or exceed weight memory during long, multi-turn sessions — which explains why a model may run smoothly at first but slow down or crash later. Mac users must also reserve 20–25% of unified memory for the operating system and display compositor before budgeting for model usage. A reliable sizing approach combines weight memory, a realistic KV cache estimate based on expected session length, and that system overhead buffer to avoid unexpected performance degradation.