Guides · Updated
Running local AI on a laptop: what actually works
What a gaming laptop, a MacBook or a plain office laptop can realistically run with Ollama, plus the laptop-specific gotchas: smaller VRAM, power limits, heat and battery.
You don’t need a desktop tower to run AI locally. Almost any recent laptop can run a small model, and a laptop with a decent GPU or plenty of unified memory can run genuinely useful ones. The catch is that laptops hide their limits in places the spec sheet doesn’t make obvious.
The number that matters: memory the model can use
- Laptop with an NVIDIA or AMD GPU: what counts is the GPU’s own VRAM, not system RAM.
- MacBook (Apple Silicon): the GPU shares unified memory, so a 32 GB MacBook can give most of that 32 GB to a model.
- Laptop with only integrated graphics: models mostly run on the CPU from system RAM. It works, but slowly.
Why memory decides everything: VRAM and quantisation explained.
Laptop GPUs aren’t desktop GPUs
The same name often means less VRAM in a laptop. Check before you assume a model will fit:
| GPU name | Laptop VRAM | Desktop VRAM |
|---|---|---|
| RTX 4050 | 6 GB | (no desktop card) |
| RTX 4060 / RTX 5060 | 8 GB | 8 GB |
| RTX 4070 / RTX 5070 | 8 GB | 12 GB |
| RTX 4080 | 12 GB | 16 GB |
| RTX 5070 Ti | 12 GB | 16 GB |
| RTX 4090 | 16 GB | 24 GB |
| RTX 5080 | 16 GB | 16 GB |
| RTX 5090 | 24 GB | 32 GB |
Two more differences: laptop GPUs usually have lower memory bandwidth than their desktop namesakes (so fewer tokens per second), and manufacturers set different power limits, so the same GPU can be noticeably faster in one laptop than another.
To see your VRAM on Windows, open Task Manager → Performance → GPU and look at Dedicated GPU memory, or run nvidia-smi on NVIDIA laptops.
What each kind of laptop can run
| Laptop | Realistic models | Try first |
|---|---|---|
| No GPU, 8 GB RAM | 1-3B | llama3.2:3b, gemma3:1b |
| No GPU, 16 GB+ RAM | 3-4B comfortably, 8B slowly | qwen3.5:4b, phi4-mini, gemma4:e2b |
| 6 GB GPU (RTX 4050) | 3-4B on the GPU | qwen3.5:4b, qwen3:4b |
| 8 GB GPU (RTX 4060/4070, 5060/5070) | 7-8B on the GPU | qwen3:8b, qwen2.5-coder:7b |
| 12 GB GPU (RTX 4080, 5070 Ti) | 12-14B on the GPU | gemma4:12b, qwen3:14b |
| 16 GB GPU (RTX 4090, 5080) | 14B with long context, 20B tight | gemma4:12b, gpt-oss:20b |
| 24 GB GPU (RTX 5090) | 27-31B on the GPU | qwen3.6:27b, gemma4:26b |
| MacBook, 8 GB | 3-4B | qwen3.5:4b, llama3.2:3b |
| MacBook, 16 GB | up to ~10 GB models | gemma4:12b, qwen3.5:9b |
| MacBook, 32 GB+ | ~20 GB models and up | qwen3.6:27b, gpt-oss:20b |
These are starting points. For your exact machine, use the GPU picker (choose No GPU / CPU only or Apple Silicon Mac where it applies). More detail per memory size: 8GB, 12GB, 16GB, 24GB and the Mac guide. MacBook owners should also read Local AI on a Mac with Apple Silicon; for laptops without a GPU, the tips in Local AI on a budget PC apply too.
Laptop gotchas
- Plug in. Many laptops cut GPU and CPU power on battery, so replies slow down, and a busy GPU drains the battery quickly.
- Heat. Long replies keep the GPU at full load. Fans ramp up, and thin laptops may slow down as they heat up. Fanless models like the MacBook Air throttle under sustained load. A hard, flat surface or a stand helps.
- Check it’s really using the GPU. Run
ollama ps: the PROCESSOR column should read100% GPU. On laptops with both Intel/AMD integrated graphics and an NVIDIA GPU, Ollama uses the NVIDIA GPU automatically as long as a current NVIDIA driver is installed. - Close other VRAM users. Browsers, video calls and games share the same GPU memory.
- RAM is often soldered. Many thin laptops can’t be upgraded later, so buy the memory you’ll need up front.
- Don’t buy for the NPU. “AI PC” TOPS ratings mostly describe the NPU, while popular local LLM tools like Ollama are built around the GPU and CPU. For running chat models, VRAM (or unified memory on a Mac) is the spec to pay for.
Laptops with lots of unified memory
A few Windows laptops and tablets use AMD’s Ryzen AI Max chips, which share up to 128 GB of memory with a strong integrated GPU. Like Macs, they can load models far bigger than any laptop GPU can hold, but they generate text more slowly than a large desktop card. Mixture-of-experts models suit them best. The same chips power the mini PCs in our AMD, Intel Arc and mini PC guide. How to choose one of those boxes: Choosing a mini PC for local AI.
Quick start
- Install Ollama on Windows, Mac or Linux.
- Plug in, then start small:
ollama run qwen3.5:4b. - Run
ollama ps. If it says100% GPUand replies feel quick, try the next size up.
Prefer a graphical app? LM Studio shows whether a model will fit your memory before you download it. See LM Studio vs Ollama. If something goes wrong, see When it doesn’t work.
Get the Local AI weekly
New models worth running, what fits which GPU, and setup fixes. One short email a week, free. Unsubscribe any time.
Local-LLM App Starter
A tested starter for building your own app on Ollama: chat, RAG and a GPU-fit checker. Skip the setup hours.
kits.lilaforge.com → VibeEngineBest local LLM tools, ranked
Ollama, LM Studio, Open WebUI and llama.cpp scored from public evidence, with top picks.
vibeengine.lilaforge.com →