How much RAM you actually need to run AI locally

说明与用途

用户在选择本地运行 AI 的 PC 时，往往关注价格或核心数，却忽视了内存是决定能否运行模型的关键因素。本文说明如何根据目标模型和上下文估算所需内存，并考虑 CPU/GPU 共享内存的实际可用量，从而挑选合适的配置，避免不必要的开支。

问题

Users often choose AI PCs based on price or specs without knowing the real limiting factor—memory—leading to underpowered setups.

结果

By focusing on memory rather than price or cores, users can select a machine that actually supports their desired model, reducing cost and ensuring functionality.

工作流

• Identify the target model size (small, mid-size, large).
• Consult the sizing table to determine required memory at 4‑bit quantisation.
• Estimate additional memory needed for context (conversation, system prompt, code repo).
• Check the machine’s unified‑memory architecture to see how much is available after OS overhead.
• Select a configuration (24 GB, 64 GB, 128 GB) that meets the total memory requirement.
• Compare prices across configurations to find the most cost‑effective option.

证据

about 5 GB at 4‑bit quantisation

about 15–20 GB at 4‑bit, before you add context

about 35–40 GB at 4‑bit, plus context

24 GB gets tight as soon as the operating system takes its share

The current generation runs from $2,599 to $5,999