Which Local AI Model Can Your Laptop Run?
Answer four quick questions about your laptop to get a local AI model that fits its memory, plus the exact Ollama commands to install it, test it and check it works.
Tell me about your laptop
DoneInstall Ollama
DoneAnswer the questions in step 1 to open this step. Go to step 1
Ollama is free software that downloads AI models and runs them on your own computer. Install it once, and every model in this guide works the same way.
Windows
Download the installer from ollama.com/download/windows and run it. It installs into your user folder and does not need administrator rights. Or install it from PowerShell:
irm https://ollama.com/install.ps1 | iexNeeds Windows 10 22H2 or newer and about 4 GB of disk space for Ollama, plus room for models. NVIDIA cards need driver 551.61 or newer. AMD Radeon cards work through a ROCm v7 driver or, on many cards, Vulkan.
Mac
Download Ollama from ollama.com/download/mac, open it and drag it to Applications. Or install it from Terminal:
curl -fsSL https://ollama.com/install.sh | shNeeds macOS 14 Sonoma or newer. Macs with an Apple chip run models on the GPU; Intel Macs run them on the processor only.
Linux
Run the official install script. It also sets Ollama up as a background service.
curl -fsSL https://ollama.com/install.sh | shNVIDIA cards need the NVIDIA driver (check with nvidia-smi). AMD cards need ROCm v7; see Ollama’s Linux guide.
Open a new terminal window (PowerShell on Windows, Terminal on Mac) and run:
ollama --versionYou should see a version number. If the terminal says ollama is not recognised, close it and open a new one. Already had Ollama? Update it: newer models such as Gemma 4 need a recent version.
Run your first prompt
DoneInstall Ollama and press “I installed it” in step 2 to open this step. Go to step 2
These commands use Gemma 4 E4B, the model I tested. Answer step 1 and they switch to the model that fits your laptop.
Download and start Gemma 4 E4B
ollama run gemma4:e4b --verboseThe first run downloads about 6.6 GB, so it can take a few minutes.
--verbosemakes Ollama print speed stats after each answer.Paste this test prompt at the
>>>promptExplain what a local AI model is in three short bullet points, and give one limitation.A good answer: three bullets that say it runs on your own computer and keeps your data on it, plus a real limitation, such as being slower or less capable than large cloud models.
Read the speed
When the answer finishes, Ollama prints timing stats.
eval rateis how fast the model writes, in tokens per second. Anything above about 5 is faster than most people read. The first answer after loading is slower.Check where it ran
ollama psThe PROCESSOR column shows whether the model is on the GPU, the CPU, or split between them. Type
/byeto leave the chat.
Common fixes
ollamais not recognised: close the terminal and open a new one, so it picks up the new install.- “model not found” or the download fails: check the tag is typed exactly, then update Ollama to the latest version.
- Very slow, or the computer freezes: close other apps, then try a smaller model. Run
ollama psto see whether it is on the CPU. - The graphics card is not used: update the graphics driver and Ollama, then check
ollama psagain. See Ollama’s hardware support list. - A long “Thinking…” section, or the answer stops early: some models plan before they answer. Add
--think=falseto the run command for quick replies.
Still stuck? The troubleshooting table in my full write-up covers more cases.
Make it more useful
DoneRun the test prompt in step 3 and confirm it answered to open this step. Go to step 3
Compare it with another model
Ask both the same question from your own work.
ollama run gemma4:e4bSkip the thinking for quick replies
This model plans before it answers. Turn that off when you want a fast reply:
ollama run gemma4:e4b --think=falseWork with longer documents
Below 24 GB of graphics memory, Ollama keeps about 4,000 tokens of conversation in view. For long files, raise the context length in the Ollama app’s settings. It uses more memory, so raise it in steps.
Use it inside VS Code
The free Continue extension connects VS Code to Ollama, so you can ask about your code without leaving the editor. My full write-up has the exact config.
Keep it private
Skip model tags that end in cloud: those run on Ollama’s servers, not your laptop. And keep Ollama’s port, 11434, closed to your network, because its API has no login.
Every model in this guide
Download sizes are for the default tag on Ollama’s library, checked on 17 November 2026. On a Mac, Ollama can pick a different build of the same model, so the download may be larger.
| Model | Ollama tag | Download | Accepts | Context | Good for | Comfortable with |
|---|---|---|---|---|---|---|
| Qwen 3.5 2B Alibaba | qwen3.5:2b | 2.7 GB Mac: up to 3.1 GB | Text, images | 256K | chat, images | 16 GB RAM, a 6 GB graphics card, or an 8 GB Mac |
| Gemma 4 E2B | gemma4:e2b | 4.6 GB Mac: up to 7.5 GB | Text, images | 128K | chat, images | 16 GB RAM, an 8 GB graphics card, or a 16 GB Mac |
| Qwen 3.5 4B Alibaba | qwen3.5:4b | 3.3 GB Mac: up to 4 GB | Text, images | 256K | chat, coding, images, maths and logic | 16 GB RAM, a 6 GB graphics card, or a 16 GB Mac |
| Gemma 4 E4B | gemma4:e4b | 6.6 GB Mac: up to 9.5 GB | Text, images | 128K | chat, coding, images | 24 GB RAM, a 12 GB graphics card, or a 24 GB Mac |
| Qwen 3.5 9B Alibaba | qwen3.5:9b | 6.6 GB Mac: up to 7.6 GB | Text, images | 256K | chat, coding, images, maths and logic | 24 GB RAM, a 12 GB graphics card, or a 16 GB Mac |
| Gemma 4 12B | gemma4:12b | 8 GB | Text, images | 256K | chat, coding, images, maths and logic | 24 GB RAM, a 12 GB graphics card, or a 16 GB Mac |
| gpt-oss 20B OpenAI | gpt-oss:20b | 14 GB | Text | 128K | chat, coding, maths and logic | 32 GB RAM, a 16 GB graphics card, or a 32 GB Mac |
| Gemma 4 26B | gemma4:26b | 18 GB Mac: up to 19 GB | Text, images | 256K | chat, coding, images, maths and logic | 48 GB RAM, a 24 GB+ graphics card, or a 48 GB Mac |
| Qwen 3.8 27B Alibaba | qwen3.8:27b | 18 GB | Text, images | 256K | chat, coding, images, maths and logic | 48 GB RAM, a 24 GB+ graphics card, or a 48 GB Mac |
How the memory check works
Each model needs roughly its download size plus about 1.5 GB for a short conversation. The finder compares that with what your laptop can spare:
- No separate graphics card: up to half your RAM counts as a comfortable fit, and up to three quarters if you close other apps.
- NVIDIA or AMD graphics card: a model that fits in the card’s memory runs fastest. If it does not fit, Ollama splits it between the card and your RAM, which is slower.
- Mac with an Apple chip: the processor and GPU share memory, and macOS keeps part of it for itself, so about 60% counts as comfortable.
Speed labels are estimates. Real speed depends on your processor, memory and drivers, so check the eval rate Ollama reports on your own machine.
Common questions
Can I run a local AI model without a graphics card?
Yes. Ollama runs models on the processor and normal RAM. On my laptop, with integrated graphics and 32 GB of RAM, Gemma 4 E4B wrote about 25.7 tokens per second, which is faster than most people read.
How much RAM do I need for a local LLM?
As a rough guide: 8 GB runs only the smallest models, such as Qwen 3.5 2B, with other apps closed. 16 GB handles 4B-class models comfortably and 9B-class models with other apps closed. 32 GB runs gpt-oss 20B. Larger models such as Qwen 3.8 27B want 48 GB of RAM or a 24 GB graphics card.
Is running AI locally free and private?
Ollama is free, and the models here are free to download under their own licences. Your prompts stay on your computer as long as you avoid model tags ending in cloud, which run on Ollama’s servers, and keep Ollama’s port closed to your network.
Gemma 4 or Qwen 3.5: which should I pick?
Both are good at laptop sizes. On my laptop, Gemma 4 E4B was faster (25.7 against 16.1 tokens per second) and answered straight away, while Qwen 3.5 thinks before it answers, which can help on harder questions. Try both on a question from your own work.
Sources
- Ollama model library: Gemma 4, Qwen 3.5, Qwen 3.8 and gpt-oss
- Ollama documentation: Windows, macOS, Linux, hardware support and context length
- My tests: Run local LLMs on a Windows 11 laptop with Ollama