How much RAM do I need for local LLMs?

16 GB runs the 7–8B workhorses, 32 GB unlocks 13B comfortably and 30B at a pinch, and 64 GB is the entry ticket to 70B-class models. Q4 weights need roughly 5 GB for 7B, 9 GB for 13B, 20 GB for 30B and 43 GB for 70B, plus 4–8 GB for the OS and context.

Does dual-channel RAM matter for AI?

A great deal. CPU inference speed is limited almost entirely by memory bandwidth, so a single 16 GB stick delivers close to half the tokens per second of 2×8 GB. When buying used, verify both slots are populated in HWiNFO.

Does more RAM make models run faster?

No. More RAM runs bigger models, not faster ones — 64 GB does not speed up a small model versus 32 GB. Buy capacity for the model sizes you want and channels, meaning two sticks, for the speed you want.

Is DDR5 noticeably better than DDR4 for local LLMs?

Yes, visibly. DDR5-4800 is a big part of why a ThinkPad T14 Gen 3 manages 4–5 tok/s on a 7B model where older DDR4-2666 machines manage about 2.5–3 tok/s at the same size.

Can a cheap laptop run a 70B model?

Yes — 64 GB of RAM in a roughly £300 Dell Latitude 5540 runs 70B models, slowly at about 1–2 tok/s. SO-DIMM upgrades of about £40–£130 are the best pound-per-capability spend in local AI; soldered-RAM laptops forfeit that path.

How Much RAM Do You Need for Local LLMs? 16 vs 32 vs 64 GB Explained

VRAM gets all the attention, but system RAM quietly decides what your laptop can run at all — especially on the integrated-graphics machines that dominate the under-£500 market. This guide gives you the sizing maths, explains why how your RAM is installed matters as much as how much, and shows which upgrades pay off.

What RAM actually does in local AI

Three jobs. First, on CPU-only machines the entire model must fit in RAM — there is no VRAM to hide in. Second, on GPU machines RAM is the overflow pool: llama.cpp and Ollama split models between VRAM and RAM, so a 6 GB card plus 32 GB RAM runs models the card alone never could. Third, context (the KV cache) grows with your conversation and lands wherever the model lives — long-context work adds gigabytes on top of the model size.

The sizing table

Approximate memory needed just for the model weights at Q4_K_M quantisation (see our GGUF quantisation guide for what Q4 means) — add 4–8 GB for Windows, your browser, and context:

Model size	Q4 weights	Comfortable on CPU with
3–4B	~2–2.5 GB	8 GB (16 GB realistically)
7–8B	~4–5 GB	16 GB
13–14B	~8–9 GB	16 GB tight, 32 GB comfortable
30–34B	~18–20 GB	32 GB tight, 64 GB comfortable
70B	~40–43 GB	64 GB only

So the practical tiers: 16 GB runs the 7–8B workhorses; 32 GB unlocks 13B comfortably and 30B at a pinch; 64 GB is the entry ticket to 70B-class models on a laptop — slowly (roughly 1–2 tok/s on the CPUs we review, estimated), but they run, which is remarkable for a £350 machine like a maxed-out Dell Latitude 5540.

Speed: bandwidth beats capacity

Here’s the part most buyers miss. CPU inference speed is limited almost entirely by memory bandwidth — every generated token reads the whole model from RAM. Consequences:

Dual-channel is not optional. One 16 GB stick gives you half the bandwidth of 2×8 GB — and close to half the tokens per second. When buying used, verify both slots are populated (HWiNFO → Memory → channel count); single-stick configs are common in off-lease machines and silently cripple LLM work.
DDR5 beats DDR4 visibly. The DDR5-4800 in a ThinkPad T14 Gen 3 is a big part of why it manages 4–5 tok/s on a 7B where older DDR4-2666 machines manage ~2.5–3.
More RAM doesn’t mean faster. 64 GB runs bigger models than 32 GB; it doesn’t run small models any quicker. Buy capacity for the models you want, channels for the speed you want.

On GPU machines the same logic explains offloading behaviour: every layer that spills from VRAM to RAM drops from ~450 GB/s (GDDR6) to ~50–75 GB/s (laptop DDR4/5). A 13B split 80/20 between a 6 GB card and RAM still feels fine; split it 50/50 and you’re mostly CPU-bound. That’s why we bang on about VRAM so much — but RAM decides whether the spillover works at all.

Upgrade economics (July 2026, estimates — check current prices)

SO-DIMM DDR4 is the cheapest capability upgrade in this hobby: roughly £40–£60 for 2×16 GB, £90–£130 for 2×32 GB. DDR5 runs higher (~£120–£170 for 2×32 GB). Against that:

Dell Latitude 5540: to 64 GB — the cheapest 70B-capable path we know: ~£280 laptop + ~£110 RAM.
Legion 5 Gen 6/Gen 7: to 32–64 GB — pairs a 6–8 GB GPU with a deep offload pool; the best all-round value combo on the site.
Precision 7560 / ThinkPad P15 Gen 2: four slots, up to 128 GB — overkill for most, priceless for 70B experiments alongside 16 GB VRAM.
The warning: ThinkPad T14s Gen 4 and most thin-and-lights solder their RAM. What you buy is what you die with — buy the 32 GB config or don’t.

Quick recommendations

Buying under £400: prioritise two RAM slots over almost everything else. 16 GB dual-channel today, 32–64 GB for £50–£110 whenever ambition strikes.
Buying a GPU laptop: 16 GB RAM minimum, 32 GB if 13B+ models are the plan. RAM should be ~2× your VRAM as a lazy rule of thumb.
Already own it: check channel config today — populating an empty second slot is a £25, five-minute upgrade that can nearly double your tok/s.

Summary

Q4 model weights need roughly: 7B → 5 GB, 13B → 9 GB, 30B → 20 GB, 70B → 43 GB — plus 4–8 GB for the OS and context. 16/32/64 GB are the real capability tiers.
CPU inference speed = memory bandwidth: dual-channel roughly doubles tok/s versus a single stick, and DDR5 visibly beats DDR4.
On GPU machines, RAM is the offload pool — it decides whether models bigger than your VRAM run at all.
SO-DIMM upgrades (~£40–£130) are the best pound-per-capability spend in local AI; soldered-RAM laptops forfeit it, so spec them right on day one.
64 GB in a £300 Latitude runs 70B models. Slowly. But it runs them.