AISignal
Subscribe

How Much RAM for Local LLMs and AI Development (2026): Buying Guide

System RAM matters for local AI in two ways: it holds whatever doesn't fit in GPU memory, and it runs everything else, from containers to vector databases. If you run models on the CPU or offload layers from a small graphics card, RAM capacity decides what fits.

Picks are grouped by workload and by the kind of memory your machine takes.

Last updated · How we picked

As an Amazon Associate we earn from qualifying purchases. Links open an Amazon search by spec, where Amazon shows current models, prices and availability. We list no prices, ratings or discounts.

Check how much memory a model needs in the run-locally calculator →

Quick picks by situation

  1. Best for general AI development

    32 GB DDR5 memory kit

    Who it's for
    Developers running containers, an IDE, a browser and cloud-model tools.
    Why
    32 GB covers a typical development stack with room for a small model on the CPU.
    Skip if
    Your machine takes DDR4 or laptop-size SO-DIMM memory.
    What to check
    Confirm your board takes DDR5 desktop DIMMs, its supported speeds, and whether you are adding to or replacing existing sticks.
    See options on Amazon

  2. Best for CPU offload and mid-size models

    64 GB DDR5 memory kit

    Who it's for
    People with a small graphics card who offload layers, or who keep several services in memory.
    Why
    64 GB holds a 32B model at 8-bit (about 36 GB by our estimate) alongside the rest of the system.
    Skip if
    Your models already fit in your graphics card's VRAM.
    What to check
    Two matched sticks are simpler than mixing kits. Check the motherboard's maximum capacity per slot.
    See options on Amazon

  3. Best for 70B models on the CPU

    128 GB DDR5 memory kit

    Who it's for
    People who want to load 70B-class models without a large graphics card.
    Why
    A 70B model needs about 41 GB at 4-bit and about 76 GB at 8-bit by our estimate, so 128 GB leaves room for the system and other work.
    Skip if
    You need fast responses; CPU generation speed is limited by memory bandwidth.
    What to check
    Four-stick kits often run at lower speeds than two-stick kits; check the board maker's memory support list.
    See options on Amazon

  4. Best for older desktops and servers

    64 GB DDR4 memory kit

    Who it's for
    Owners of DDR4 desktops and used servers.
    Why
    Many homelab boxes and older workstations take DDR4, and more of it extends their useful life for containers and small models.
    Skip if
    Your board takes DDR5; the two are not interchangeable.
    What to check
    Check the generation, the board's maximum capacity, and whether a server needs registered (RDIMM) or ECC memory.
    See options on Amazon

  5. Best for mini PCs and laptops with memory slots

    64 GB DDR5 SO-DIMM kit

    Who it's for
    Owners of mini PCs and laptops with SO-DIMM slots.
    Why
    A larger SO-DIMM kit is often the simplest upgrade for a small machine running containers or small models.
    Skip if
    Your machine has soldered memory, as many thin laptops and unified-memory mini PCs do.
    What to check
    Confirm the machine has SO-DIMM slots, takes DDR5, and supports the kit's capacity.
    See options on Amazon

How to choose

DDR4 or DDR5, DIMM or SO-DIMM

Memory must match the machine: generation (DDR4 or DDR5), size (desktop DIMM or laptop SO-DIMM) and the board's maximum capacity. Check the motherboard or system maker's spec page first.

Size by workload

32 GB suits most development. 64 GB gives room for CPU offload of mid-size models and several services. 128 GB or more is for loading 70B-class models on the CPU or running many VMs.

Two sticks or four

Two larger sticks usually run at higher rated speeds than four smaller ones on consumer boards, and leave slots free for later. Buy a matched kit rather than mixing.

System RAM is not VRAM

Adding system RAM lets a larger model load, but layers on the CPU run slower than on a graphics card. Unified-memory Macs and mini PCs share one pool and can't be upgraded after purchase.

FAQ

How much RAM do I need to run a local LLM?

If the model runs on a graphics card, 32 GB of system RAM is usually plenty. If it runs on the CPU, you need roughly the model's size plus room for the system: about 6 GB for a 7B model at 4-bit, about 41 GB for a 70B model.

Does more RAM make models run faster?

Not by itself. More RAM lets a larger model load; speed depends on whether it runs on the GPU and on memory bandwidth.

Can I mix memory kits?

It often works but can force lower speeds or cause instability. A matched kit of the capacity you need is the safer choice.

How we picked

  • We grouped memory by workload and by the kind of slot it needs, since a kit that doesn't match the board is the most common mistake.
  • Model-size fits come from the same estimate as our run-locally calculator (4-bit weights, 8K context, about 75% of unified memory usable by the GPU). Real use varies by runtime, context length and model.
  • Links open Amazon searches by spec, not a single listing, so you compare current models, prices and reviews yourself.
  • We have not tested these products hands-on for this guide and make no claims about speed or how any listing performs.

Related buying guides

More: All buying guides · Black Friday guide · Deals