How much RAM do I need for local LLMs?

If the model fits in GPU VRAM, 32GB system RAM is often enough. For serious offload or 30B–70B work, 64GB DDR5 is the default; 128GB+ for heavy CPU offload or multi-model workflows.

Does system RAM matter if I have a strong GPU?

Yes for the OS, context, and any layers that spill out of VRAM. When the whole model stays on-GPU, RAM capacity matters less than dual-channel stability — but undersizing still causes swapping.

What DDR5 speed is best for local AI?

On AMD AM5, DDR5-6000 CL30 EXPO dual-channel is the practical pick. On modern Intel, 6000–6400 UDIMM is the sensible band. Capacity usually beats extreme MT/s.

Should I use four sticks to reach 128GB for AI?

Prefer two high-capacity sticks (2×64GB) when your board supports them. Four-stick DDR5 setups often force lower stable speeds and harder training.

Best RAM for Local AI / LLM Inference (2026)

Local LLMs, image models, and agent stacks live or die on memory. The most common mistake is treating system RAM and GPU VRAM as the same pool.

Bottom line up front: Size GPU VRAM to the model first. Then buy fast dual-channel DDR5 for everything that does not fit on the GPU — OS, context, KV cache overflow, and CPU offload. For most serious local-AI desktops in 2026: 64GB DDR5-6000 (2×32GB). Use 32GB only for small models that mostly fit in VRAM. Use 128GB+ when you routinely offload 70B-class weights or run multi-model workflows.

Compare live kits on DDR5 RAM Prices and Best RAM Deals. For capacity planning across all workloads, also see How Much RAM Do I Need?.

Quick decision table

Your setup — System RAM target — Why
Your setup	System RAM target	Why
7B–8B Q4 mostly in VRAM	32GB DDR5	OS + browser + light headroom
13B–34B with some offload	64GB DDR5	Offload layers + long context
70B Q4 with GPU offload	64–128GB DDR5	Weights spill + KV cache
CPU-only / heavy offload 70B	128GB+ DDR5	Entire weight path hits DRAM
Apple Silicon unified memory	Buy max at purchase	Not upgradeable later

VRAM vs system RAM (the rule that saves money)

1If the model fits in VRAM (or unified memory), token speed is dominated by the GPU/accelerator. System RAM still needs to run Windows/macOS, the UI, and context — but buying 128GB “just because AI” is often wasted.
2If layers offload to CPU RAM (llama.cpp, Ollama, LM Studio, etc.), system capacity and DDR5 bandwidth become the bottleneck. Slow or single-channel memory feels like the model is broken.
3Quantization (Q4 / Q5 / Q8) shrinks weights. A 70B Q4 package is roughly ~40GB of weights — not 140GB FP16 — but you still need headroom above that number.

Rough weight footprints (order of magnitude):

Model class — Q4 weights (approx.) — Comfortable system RAM when offloading
Model class	Q4 weights (approx.)	Comfortable system RAM when offloading
7B–8B	4–6GB	32GB
13B–14B	7–10GB	32–64GB
30B–34B	18–22GB	64GB
70B–72B	35–45GB	64–128GB

Always add OS overhead (often 8–16GB in real desktops) and context growth.

Why DDR5 speed still matters for AI

CPU-side and offload inference stream weights from memory every token. Higher sustained bandwidth improves tokens/sec when the GPU is not doing all the work.

Platform — Practical AI memory pick
Platform	Practical AI memory pick
AMD AM5	DDR5-6000 CL30 EXPO, dual-channel — see Best RAM for AM5
Intel LGA1851	DDR5-6000–6400 UDIMM; CUDIMM only if chasing extreme XMP — CUDIMM Guide
Older DDR4	Upgrade capacity first; bandwidth lags DDR5 for offload

Do not buy four sticks to hit 128GB on a daisy-chain board unless you accept lower stable speeds — Four Sticks DDR5 Stability. Prefer 2×64GB or 2×48GB when the board/CPU allow it.

Recommended builds (commerce-first)

Starter local AI (chat + coding assistants)

32GB (2×16GB) DDR5-6000 CL30
GPU with enough VRAM for your main 7B–14B Q4 model
Good for: Ollama daily drivers, Continue.dev, small agents

Shop: DDR5 prices · Best deals

Serious local AI (default recommendation)

64GB (2×32GB) DDR5-6000 CL30–CL32
Headroom for 30B-class offload, longer context, browser + IDE + model UI
Best $/usefulness point for most builders in 2026

Heavy offload / research

128GB (2×64GB) DDR5-5600–6000 (stability over peak MT/s)
Or workstation ECC if uptime/data integrity matters — Server / ECC Guide
Validate QVL; enable XMP/EXPO carefully — XMP/EXPO BIOS Guide

Laptops

Prefer machines with upgradeable SODIMM or LPCAMM2 if you expect model growth — Laptop RAM Upgrade and LPCAMM2 / CAMM2 Guide.
Soldered LPDDR means you must buy capacity at purchase.

Stability and verification

Local AI stresses memory for hours. After enabling XMP/EXPO:

1Confirm dual-channel in CPU-Z.
2Run a memory stress pass before large downloads fill your return window — How to Test RAM Health.
3Avoid mixing random used sticks for a 128GB Frankenstein kit — Buying Used RAM and Mixing RAM.

What not to optimize

Chasing 8000+ MT/s CUDIMM on AM5 for AI — bypass mode wastes money (CUDIMM Guide).
RGB premiums — irrelevant to tokens/sec.
Waiting for DDR6 — consumer platforms are not a 2026 buy reason (DDR5 vs DDR6).
Ignoring VRAM — no amount of DDR5 fixes a model that needed 24GB VRAM and you have 8GB.

Buying checklist

1List the largest model + quantization you will actually run.
2Check whether it fits in VRAM; if not, size RAM for offload + OS.
3Buy a matched dual-channel DDR5 kit (32 / 64 / 128GB).
4Prefer 6000 CL30-class on AM5; solid 6000–6400 on modern Intel.
5Enable EXPO/XMP; stress test; then load models.
6Track market timing on RAM Price Trends if the upgrade is optional.

Capacity and homelab paths

Compare Best 96GB DDR5 when 64GB is too small. If the same host runs virtualized services, use the Proxmox RAM guide to budget guests and storage headroom.

Related guides

Live kits: Best 64GB DDR5 · DDR5-6000 CL30
How Much RAM Do I Need?
Best RAM for AM5
RAM Buying Guide
Best RAM for Gaming (if the same PC also games)