How much RAM do I need to run an AI model locally?

It depends on the model's size. A 4B-parameter model needs roughly 4GB of RAM at Q4 quantization, a 9B model about 7GB, and a 70B model around 45GB. With 16GB of RAM you can run most models up to about 12B parameters comfortably. The smallest model we track needs 808 MB.

Do I need a graphics card to run open-source AI models?

No. Models run on ordinary system RAM using tools like Ollama or LM Studio, and Apple Silicon Macs are particularly good at it because they share memory between CPU and GPU. A dedicated graphics card is faster, but its VRAM is a hard ceiling — a 24GB card cannot run a model needing 30GB, even if the machine has 64GB of system RAM.

Are open-source AI models free to use?

The weights are free to download and run on your own hardware, so there is no per-token cost. Licences differ though — some permit commercial use freely, others restrict it. Each model page lists its licence.

Open-source AI models

Open-source models are free to download and run on your own computer — no API key, no monthly fee, and nothing sent to anyone else's server. The catch is memory: a model has to fit in your RAM. Of the 51 language models we track, 29 run on ordinary consumer hardware.

What can your computer run?

Two questions: how much memory you have, and what kind. The first decides what loads, the second decides how fast it runs.

How much memory do you have?

System RAM

Graphics card

— a hard ceiling; a model can't spill into system RAM

What kind of machine is it?

Desktop / replaceable memory (DDR)

Laptop / soldered memory (LPDDR)

Mac — Apple Silicon

Numbers beside each are memory bandwidth in GB/s — the figure that sets speed.

Which Mac is which: plain M-chips are the MacBook Air, Mac mini and iMac · Pro and Max are the MacBook Pro and Mac Studio · Ultra is Mac Studio only. Apple shares one fast memory pool between CPU and GPU, which is why a Mac runs far larger models than a PC laptop with the same RAM.

On a Desktop or laptop — DDR5-5600 with 16GB RAM, 13 of 51 models fit. Five of the most-run, smallest first:

Full list for 16GB RAM →

Memory figures are Q4_K_M at 4K context, including the context cache and runtime overhead. Speed is calculated, not measured: 90 GB/s × 0.65 ÷ bytes read per token.

Bandwidth source: 5,600 MT/s × 128-bit dual channel ÷ 8 = 89.6 GB/s. How these are worked out →

Browse by what it takes to run

Sorted by hardware, not by benchmark score.

Runs on a normal laptop

Fits in 16GB of RAM. No graphics card needed.

13 models

Runs on a good laptop or desktop

Wants 32GB of RAM, or a graphics card with 24GB of VRAM.

9 models

Needs a workstation

Wants 64–128GB. Realistically a high-end Mac or a multi-GPU build.

7 models

Too big for any home computer

Hundreds of gigabytes and a rack of GPUs. See exactly what they cost to run.

22 models

Jump to a memory size

8GB 16GB 24GB 32GB 64GB 128GB

Or a graphics card

8GB VRAM 12GB VRAM 16GB VRAM 24GB VRAM 32GB VRAM

Scale

How big do these things actually get?

Kimi-K3 has 2780B parameters and needs 1575 GB of memory — about 22 H100 cards. Nobody runs it at a desk, but the numbers are worth seeing.

Most-downloaded of the models we track

HuggingFace downloads over the last 30 days, among the 51 current models listed here. Older models still in wide use — Qwen3, Llama 3, Qwen2.5 — out-download all of these but fall outside the 12-month window.

1Qwen3.8-27BAlibaba / Qwen · 28B19 GBneeds 2MiniMax-H3MiniMax · 33B64 GBneeds 3gemma-4-12B-itGoogle DeepMind · 12B9.6 GBneeds 4Kimi-K3Moonshot AI · 2780B · mixture-of-experts1575 GBneeds 5GLM-5.3-FlashZhipu AI · 321B385 GBneeds 6NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4NVIDIA · 18B · mixture-of-experts28 GBneeds 7GLM-5.3Zhipu AI · 753B360 GBneeds 8diffusiongemma-26B-A4B-itGoogle DeepMind · 26B · mixture-of-experts19 GBneeds 9MiniCPM5-1BOpenBMB · 1.1B1.3 GBneeds 10Qwen3.8-Flash-NextAlibaba / Qwen · 180B · mixture-of-experts199 GBneeds

New to running models yourself?

What does "quantized" mean?Why the same model comes in a dozen sizes.Why is a "small" model 40GB?Mixture-of-experts models are bigger than they look.Will it actually be fast?Having the RAM is not the same as having the speed.

Verified against HuggingFace on 2026-09-15Full model database (including paid APIs) →

Only models released in the last 120 days are listed, and an older generation drops off as soon as its successor ships. Specialised tools — OCR, speech recognition, translation, robotics and search embeddings — are left out: this list is general-purpose assistants, plus a few image generators.