Running a large language model on your own laptop has gone from novelty to genuinely useful, but the single number that decides what you can run is VRAM. System RAM matters too, especially for offloading, yet the model has to fit somewhere fast, and GPU memory is the fast place. Before buying, it helps to translate model sizes and quantization levels into concrete memory requirements so you are not guessing.
How model size maps to memory
Models are measured in billions of parameters, and quantization shrinks how many bytes each parameter needs. A 4-bit quantized model needs roughly half a gigabyte of memory per billion parameters, plus overhead for context. That rule of thumb lets you size hardware quickly.
| Model size | 4-bit VRAM need | Comfortable GPU VRAM | System RAM |
|---|---|---|---|
| 7-8B | ~5-6 GB | 8 GB | 16 GB |
| 13-14B | ~9-10 GB | 12 GB | 32 GB |
| 30-34B | ~19-22 GB | 24 GB or offload | 32-64 GB |
| 70B | ~40 GB | partial offload only | 64 GB+ |
When system RAM steps in
If a model does not fit entirely in VRAM, tools like llama.cpp split layers between the GPU and system RAM. This works, but every layer that lives in RAM runs far slower, so a 70B model on a laptop with 16 GB of VRAM might generate only a few tokens per second. That is why 32-64 GB of system RAM is valuable for larger models even though the GPU does the heavy lifting for the layers it holds.
Context length costs memory too
Longer prompts and conversations grow the key-value cache, which consumes additional VRAM on top of the model weights. A long 16K-32K context can add a couple of gigabytes, so leave headroom rather than sizing to the exact model footprint.
Practical laptop targets
For responsive 7-8B chat and coding assistants, an 8 GB mobile GPU is enough and pairs well with 16 GB of RAM. To run 13-14B models smoothly, aim for 12 GB of VRAM and 32 GB of RAM. If you want to experiment with 30B-class models, look for the highest-VRAM mobile GPUs (16 GB) and 64 GB of system RAM for offload. Expect these builds to run roughly $1400-2500 depending on the GPU tier.
Speed, not just fit
Tokens per second is what you feel
Fitting a model in memory is only half the goal; the other half is generation speed, measured in tokens per second. A 7-8B model fully on an 8 GB GPU might produce 30-50 tokens per second, fast enough to feel conversational. The same model split across GPU and system RAM can crawl at a few tokens per second. Keeping the whole model in VRAM is the single biggest factor in a responsive experience.
Quantization is your friend
Lower-precision quantization (4-bit and similar) shrinks a model to a fraction of its full size with only a modest quality drop for most everyday tasks. That is what makes 13-14B models practical on a 12 GB laptop GPU. If you need maximum quality, higher precision costs roughly double the memory, so plan your VRAM budget around the quantization level you actually intend to run.
FAQ
Can I run a 70B model on a laptop?
Only with heavy offloading to system RAM, and it will be slow, often a few tokens per second. For usable speed on a laptop, 13-14B at 4-bit is a far better fit for current mobile GPUs.
Does the GPU brand matter for local LLMs?
Mature tooling and broad support make current-generation mobile RTX GPUs the smoothest path today. Other GPUs can work, but you may spend more time on setup and hit occasional compatibility gaps.
Bottom line: VRAM decides the ceiling. Target 8 GB for 7-8B models, 12 GB for 13-14B, and 16 GB plus 64 GB of RAM if you want to stretch toward 30B-class models with offload.
