NowHardware guide
GPUs for running AI at home: how much video memory you need
Updated October 28, 2026 · By the SuperiNow team
If you want to run AI models on your own computer (chatbots like Llama, Mistral or Qwen, or image generators), the graphics card does the heavy lifting. And for AI, one spec matters more than all the others: video memory (VRAM).
Speed decides how fast a model answers. Video memory decides whether the model runs at all. A model that doesn't fit in VRAM either won't load or slows to a crawl as it spills into regular memory.
The rule of thumb
Most people run open models compressed to 4-bit (often called "Q4"), which shrinks them a lot with little loss in quality. At 4-bit, a model needs roughly:
- About half a gigabyte of VRAM per billion parameters, plus
- 1–3 GB extra for the conversation history (the "context") and overhead.
So a 7-billion-parameter model needs about 5–6 GB, a 14B model about 9–10 GB, and a 32B model about 20 GB.
Desktop cards by video memory
| VRAM | Cards (RTX 40 and 50 series) | What runs comfortably at 4-bit |
|---|---|---|
| 8 GB | RTX 4060, RTX 5060, 8 GB versions of the 4060 Ti and 5060 Ti | 7–8B models. Fine for trying local AI, tight for anything bigger. |
| 12 GB | RTX 4070, 4070 Super, 4070 Ti, RTX 5070 | 8B models with room for long chats; 12–14B models just fit. |
| 16 GB | RTX 4060 Ti 16 GB, 5060 Ti 16 GB, 4070 Ti Super, 4080 / 4080 Super, RTX 5070 Ti, RTX 5080 | 14B models comfortably, and image generators like Stable Diffusion XL and Flux with room to spare. The sweet spot for most people. |
| 24 GB | RTX 4090 | Models up to about 30B. |
| 32 GB | RTX 5090 | 30B-class models with long context, and the fastest image and video generation in this list. |
Big models (around 70B parameters) need roughly 40 GB or more at 4-bit. That means two graphics cards, or a Mac with lots of unified memory (see our Mac mini vs Mac Studio guide).
Laptop GPUs have less memory
The laptop version of a card often has less video memory than the desktop card with the same name. NVIDIA's RTX 50 laptop line:
| Laptop GPU | VRAM |
|---|---|
| RTX 5050, 5060, 5070 Laptop | 8 GB |
| RTX 5070 Ti Laptop | 12 GB |
| RTX 5080 Laptop | 16 GB |
| RTX 5090 Laptop | 24 GB |
So an "RTX 5070" laptop has 8 GB, while the RTX 5070 desktop card has 12 GB. Always check the VRAM in the laptop's spec sheet, not just the GPU name.
How to choose
- Just curious about local AI: a 12 GB card is enough to learn on.
- Want to use it seriously (coding helpers, document chat, image generation): 16 GB is the best value point.
- Want the biggest models on one card: 24–32 GB, or consider a Mac with 64 GB of unified memory or more.
- Buying used? An RTX 3090 also has 24 GB and is popular for local AI. Check its condition and power needs.
Don't forget the rest of the PC
- Power supply: high-end cards draw a lot. Check the card maker's recommended wattage.
- Case room and cooling: the biggest cards are long and thick. Measure first.
- System memory: 32 GB of regular RAM is a sensible minimum alongside a 16 GB card.
The short version: buy for video memory first, speed second. For most people, a 16 GB card is the sweet spot. Specs here come from NVIDIA's published figures; check the exact model's spec sheet before you buy, especially for laptops.