Future TechnologyThe VRAM tier list, and exactly which AI models your GPU can run
Key takeaways
- 7B models need about 8GB of VRAM, 13B models want 12 to 16GB, and 70B models need 35 to 48GB depending on quantisation.
- The RTX 3090, 4090 and 5090 all sit at 24GB, which makes a used 3090 the best value per gigabyte on the market.
- Apple silicon shares unified memory, so an M5 Max at 64GB runs 70B models that would otherwise need two 24GB cards.
- System RAM is the second wall: 32GB works for small models, 64GB is the real target, 128GB opens up the large ones.
8GB runs a 7B model. 48GB runs a 70B one. Almost every wasted download and failed load in local AI sits in the gap between those two numbers, because local LLM VRAM requirements get quoted as ranges rather than as the hard ceiling they actually are.
This is the GPU-side companion to the system RAM guide. Video memory is the constraint that bites first, so it gets its own tier list.
Local LLM VRAM requirements by tier
7B models need about 8GB of VRAM. That covers most laptop discrete GPUs and the entry desktop cards.
13B models want 12 to 16GB. This is where quantisation starts mattering, because a 4-bit quant of a 13B model fits in 12GB while the 8-bit version does not.
70B models need 35 to 48GB depending on quantisation. There is no consumer single card at that capacity, so this tier means either two 24GB cards or Apple silicon.
The consumer sweet spot is still 24GB
The RTX 3090, 4090 and 5090 all sit at 24GB. Three generations, one capacity, which tells its own story about how Nvidia segments the consumer line.
A used 3090 remains the best value per gigabyte on the market. It is slower than a 4090 on token generation and it draws more power for the result, but capacity determines what loads at all, and speed only determines how long you wait. Given what memory prices are doing this year, the used market is doing more work than usual.
Apple silicon changes the arithmetic
Unified memory is shared between CPU and GPU, so the tier boundaries move. An M5 Pro at 16 to 32GB handles 7B to 14B models comfortably. An M5 Max at 64GB or more runs 70B models that would otherwise need two 24GB cards.
It is slower than a pair of discrete cards. It is also one machine, on one power supply, that does not need a case rebuild or a 1000W PSU.
System RAM is the second wall
VRAM decides what loads. System RAM decides whether the load succeeds and whether anything else on the machine stays usable while it runs.
32GB works for small models. 64GB is the real target and the point at which you stop thinking about it. 128GB opens up the large models and any serious attempt at layer offloading. A 32GB DDR5 kit is the cheapest meaningful upgrade most people can make here, and it is getting more expensive every quarter.
Which model to run at each tier
Qwen 3.5 7B is the best thing you can run on a normal laptop, with switchable fast and reasoning modes.
Gemma 3 27B for anything with 16GB or more. It is the tier where local output starts feeling close enough to a hosted model that the trip stops being worth it.
DeepSeek R1 Distilled 32B when you want chain of thought for maths and multi-step code.
Llama 3 8B remains the safe default when you want something that just works.
What to watch
Subscription prices are not falling and hosted model pricing keeps moving, which is most of the case for running locally in the first place. The tier that matters is the one your existing card already sits in. Check that number before downloading anything, because a 70B download on a 12GB card is forty gigabytes of disk you will delete.
Some links in this article are affiliate links. We may earn a small commission at no extra cost to you.