Can I run AI with 8 GB of memory?
What fits on an 8 GB graphics card, an 8 GB Mac and an 8 GB PC, and what to do when it does not.
Short answer
~8B
Yes, with small models: up to about 8B parameters in Q4 on an 8 GB graphics card, with no margin.
- On an 8 GB Mac or a PC without a graphics card, the estimate stops near 8B.
- Conversations: Keep them short; long chats need more memory.
- Bigger models: They run with part in system RAM, much more slowly.
Planning estimate with the site's formula, not a measurement. Methodology
Size by size
Q4 file, short conversations. Estimates rounded up to whole GB.
Each estimate includes a reserve: a graphics card keeps 2 GB aside for its runtime and display, because the system runs from the PC's own RAM; a Mac or a PC without a card shares its memory with the system, so the site keeps 25% of it aside, at least 2 GB: 2 GB on 8 GB (VRAM and unified memory).
| Model size | Q4 weights, est. (GB) | Needs on an 8 GB graphics card (GB) | Needs on an 8 GB Mac or PC (GB) |
|---|---|---|---|
| 1B | 0.6 | 3 | 3 |
| 3B | 1.9 | 5 | 5 |
| 4B | 2.5 | 5 | 5 |
| 8B | 5 | 8 | 8 |
| 9B | 5.6 | 9 | 9 |
| 12B | 7.5 | 11 | 11 |
| 14B | 8.8 | 13 | 13 |
8 GB machines
The largest common model size whose Q4 estimate stays within what each machine offers.
- RTX 4060 · 8 GB VRAM
- Up to 8B in Q4Offers 8 GB to a model
- RTX 5060 · 8 GB VRAM
- Up to 8B in Q4Offers 8 GB to a model
- Radeon RX 9060 XT · 8 GB VRAM
- Up to 8B in Q4Offers 8 GB to a model
- Mac mini M2 · 8 GB unified
- Up to 8B in Q4Offers 8 GB to a model
- MacBook Neo 13-inch · A18 Pro · 6-core CPU / 5-core GPU · 8 GB unified
- Up to 8B in Q4Offers 8 GB to a model
A real 9B file
The Q4_K_M file of Qwen3.5 9B is 5.7 GB (unsloth/Qwen3.5-9B-GGUF, read Oct 17, 2026). It looks like it fits an 8 GB card, but the program needs working room on top of the file: the site's estimate is 5.6 GB of weights, 1.1 GB of working room and 2 GB of reserve, about 9 GB in total.
That is more than 8 GB, so by the site's estimate a 9B model in Q4 does not fit an 8 GB card. The options are below: a smaller file, a smaller model, or splitting the work with system RAM. Qwen3.5 9B
When it is too big
- Pick a smaller file
- Q3 and Q2 files trade more quality for sizellama.cpp measures Q3_K_M at 4 and Q2_K at 3.16 bits per weight, against 4.89 for Q4_K_M.llama.cpp · quantize
- Pick a smaller model
- A 4B model leaves room for a longer conversationIn Q4 it needs about 5 GB on a graphics card, 3 GB less than an 8 GB card offers.
- Split it with system RAM
- llama.cpp and the apps built on it can run the rest on the processorIt works, much more slowly; hardware pages list these models apart, under “Also runs slowly”.llama.cpp README
- Keep conversations short
- Long chats and documents need extra memoryThe working room covers short use only.