← All articles
2026-07-027 min read

How Much RAM Do You Need for Local AI?

Practical RAM guidelines for local AI in 2026: what runs on 8, 16, 32 and 64 GB, how quantization shrinks models, and how to pick a model that fits.

The quick answer: 8 GB gets you started with small models, 16 GB is the comfortable sweet spot for good everyday AI, 32 GB opens up mid-size models and image generation side by side, and 64 GB+ is enthusiast territory for large models. RAM is the single most important spec for local AI (more than CPU speed, more than almost anything else) because the entire model has to fit in memory to run. Here's how to translate your machine's RAM into what you can actually do with it.

Why RAM is the bottleneck

A language model is, at its core, a very large table of numbers ("weights"). To generate text, your computer has to read essentially all of those numbers for every word it produces. That only works at usable speed if the whole model sits in fast memory.

So the question "can my computer run this model?" mostly reduces to: is the model file smaller than the memory I can spare? Leave a few gigabytes free for your operating system and the model's working space.

Two hardware notes before the numbers:

  • Apple Silicon Macs use unified memory: CPU and GPU share one pool, so all of your RAM is available to models. This is why Macs punch above their weight in local AI.
  • Windows/Linux PCs with a discrete GPU care about VRAM (the graphics card's own memory) for full-speed inference. A model that doesn't fit in VRAM can spill into system RAM, which works but is slower. The same size rules of thumb apply; just apply them to your VRAM first.

Quantization: why the numbers are smaller than you'd think

Models are described by parameter count (3B, 8B, 70B), where "B" means billions of weights. Stored at full precision, one weight takes 2 bytes, so an 8B model would need ~16 GB before you've opened a browser tab.

Quantization fixes this. It rounds the weights down to lower precision (commonly 4-bit), shrinking the file to roughly a quarter of full size with only a modest quality cost. Quantized models are usually distributed as GGUF files (the standard format for local AI) or as MLX versions for Apple Silicon. In practice, almost everyone runs 4-bit models, and the rules of thumb below assume that.

Rough sizes at 4-bit:

ParametersRoughly, at 4-bit
3 to 4B2 to 3 GB
7 to 8B5 to 6 GB
12 to 14B8 to 10 GB
around 30B18 to 20 GB
around 70B40 GB and up

Add a couple of gigabytes of headroom for the model's context window (its short-term memory of your conversation) and whatever else your computer is doing.

What each RAM tier gets you

8 GB: the entry ticket

Eight gigabytes holds models below the 9B class. They summarize, answer quick questions and draft short text, and they come apart on anything nuanced or multi-step. That is not a quantization problem you can fix with a better file; it is what the parameter count buys. Close your browser's 40 tabs before generating and it will at least feel faster.

Worth being blunt about the consequence: none of the models we would actually recommend fit here. If 8 GB is what you have and you want a model that holds up, the realistic route is a large one running in a browser tab rather than a small one running on your disk. LU Labs Cloud covers that, and the whole chat catalog is on every plan and every credit pack.

Verdict: fine for finding out whether local AI suits you, limiting as a daily driver.

16 GB: the sweet spot

This is where local AI starts feeling good. Models in the 7 to 14B class fit comfortably at 4-bit, and they handle the bulk of everyday tasks: solid writing, decent coding assistance, coherent long conversations. On a PC, 16 GB of VRAM is also where local image generation stops being cramped: SDXL checkpoints want 6 to 8 GB and FLUX 2 Klein wants 8 to 10, so a chat model and an image model can share the card if neither is at the top of its range.

Verdict: the recommended minimum if you're buying a machine with local AI in mind.

32 GB: room to breathe

This is the tier where the models we recommend start to fit. Qwen 3.8 27B at its default 4-bit build is 16.5 GB of weights before the context window takes its share, and our guide to running it has the full size table. Gemma 4 26B is a mixture of experts, 25.2B parameters with 3.8B active per token, so it reads like a 26B model and runs closer to the speed of a small one. Qwen3-Coder 30B, 30.5B with 3.3B active and a 262,144 token context, is the one to point a coding agent at. This tier suits people who use local AI seriously every day: heavier coding-agent work, bigger context windows, image generation as a habit rather than a party trick.

Verdict: the enthusiast sweet spot: most of the capability, none of the exotic pricing.

64 GB and beyond: the deep end

Models around 30B run easily, and 70B-class models become possible at heavier quantization. This is where local models start seriously closing the gap to cloud services on difficult tasks. It is also where local video generation gets comfortable: Wan 2.2 TI2V 5B pulls about 17 GB of files and wants 12 GB of VRAM or more, and LTX Video 2.3 wants 16 GB and up.

Verdict: wonderful if you have it; unnecessary for most people.

Practical tips whatever your tier

Don't run maxed out. If a model barely fits, your system will crawl as everything else fights for scraps. A smaller model with headroom beats a bigger one that's suffocating.

One heavy model at a time. Chat model or image model: unload one before loading the other on smaller machines.

Context costs memory too. Very long conversations and big documents inflate memory use beyond the model file itself. If things slow down in a marathon session, start a fresh chat.

Let software do the math. This is honestly the easiest fix for the whole topic: LU Labs, a free desktop AI studio, checks your actual hardware and only recommends models that fit it, across its own LU Engine, Ollama and LM Studio. No spreadsheet of quantization sizes required; see what's included. On a Mac the app is a beta for Apple Silicon on macOS 14 or newer; LM Studio is the other easy answer there, and it tells you whether a build fits before you spend the download.

Should you upgrade your RAM for local AI?

  • On a PC, often yes: system RAM is cheap to add, and a used mid-range GPU with decent VRAM transforms what you can run.
  • On a Mac, RAM is fixed at purchase, so this becomes a buying decision: if local AI is on your list, 16 GB is the floor and 32 GB is the comfortable choice. Our Mac local AI guide goes deeper on the Apple side.

But before spending anything: try what you have. Even a machine below this list's floor tells you within an evening whether local AI suits how you work, and it costs nothing but a download to find out.

There is also a route that skips the memory question entirely. The trillion-parameter models nobody runs at home are reachable in a browser on someone else's GPUs: run Qwen 3.8 without a GPU is the version of this article for the models your RAM will never hold.

The bottom line

Match the model to the memory: small models for 8 GB, 7 to 14B for 16 GB, the 26 to 30B class for 32 GB, and the big leagues from 64 GB up, all assuming the standard 4-bit quantization everyone actually uses. Pick one tier below your maximum for comfort, and you'll wonder why local AI ever had a reputation for being complicated. Questions about specific setups? The FAQ covers the common ones.

Sources