AI LUKE.
EN
Open source · Guide & explainer

Which local AI model fits your laptop?

Start with a model that leaves your computer usable. The largest model you can load is not necessarily the best choice for the work you want to do.

Luke Byrne
Luke ByrneAI Luke · Luke Byrne | AI & Tech
Video published 18 Nov 2026· 6 min read

Video in English

Watch on YouTube ↗

Choose a starting point

These are the choices in the video’s hardware table, not measured speed results for every machine. Download size is only part of runtime memory: the context, software and other applications need room too. Q4 denotes a quantized version; a different quantization changes the memory requirement.

Your hardware: Intel Mac · 8–16 GB RAM · Starting choice: Qwen3 1.7B · Larger option / extra room: Qwen3 4B; 8B on 16 GB is another candidate · What to check: CPU only in Ollama. Memory fit does not establish useful speed.

Your hardware: M-series Mac · 8 GB · Starting choice: Qwen3 1.7B · Larger option / extra room: Qwen3 4B Q4 · What to check: Shared with macOS and apps.

Your hardware: M-series Mac · 16 GB · Starting choice: Dolphin 3 8B Q4 · Larger option / extra room: Qwen3 14B Q4 · 9.3 GB download · What to check: Start at 4K context, close heavy apps and check the loaded model.

Your hardware: M-series Mac · 24 GB · Starting choice: Qwen3 14B Q4 · Larger option / extra room: More room for context and apps · What to check: 24 GB and 32 GB are different steps.

Your hardware: M-series Mac · 32–48 GB · Starting choice: Qwen3 14B Q4 · Larger option / extra room: Qwen3 32B Q4 · 20 GB download · What to check: 48 GB gives the same 32B model more headroom.

Your hardware: Windows · integrated graphics · Starting choice: 8 GB RAM: Qwen3 1.7B · Larger option / extra room: 4B; with 16 GB RAM, also try 8B · What to check: GPU support varies. Check whether Ollama falls back to CPU.

Your hardware: Windows · 4 GB dedicated VRAM · Starting choice: Qwen3 1.7B · Larger option / extra room: Qwen3 4B Q4 if it fits · What to check: Leave room for runtime and context.

Your hardware: Windows · 8 GB dedicated VRAM · Starting choice: Dolphin 3 8B Q4 · Larger option / extra room: Keep the 8B tier as the starting choice · What to check: An RTX 4060 Laptop GPU has 8 GB. System RAM is separate.

Your hardware: Windows · 12–16 GB dedicated VRAM · Starting choice: 8B Q4 · Larger option / extra room: Qwen3 14B Q4 · What to check: 12 GB is tighter; check full GPU placement at 4K context.

Your hardware: Windows · 24 GB dedicated VRAM · Starting choice: Qwen3 14B Q4 · Larger option / extra room: Qwen3 32B Q4 · What to check: Check context and GPU placement before expanding the input.

For a 16 GB M4 Mac, an 8B model is the easier everyday starting point; Qwen3 14B Q4 is a larger candidate to try. Compare the answer and waiting time on your task, alongside memory pressure. A larger parameter count does not guarantee a better answer.

Install and run Ollama

Download the Mac or Windows installer from Ollama (https://ollama.com/download), then install it normally. Open Terminal on Mac or PowerShell on Windows.

The command shown in the video is:

ollama run dolphin3:8b

The first run downloads the model. Subsequent runs reuse the saved download. If the hardware table points you to another model, use its exact tag from the Ollama library rather than assuming every model uses the same name.

In a second terminal window, inspect the running model:

ollama ps

  1. Check the PROCESSOR column. A model can run on GPU, CPU or a mixture. Memory capacity alone does not tell you the waiting time. Compare with your usual applications open, then reduce the workload if memory pressure is high.
  2. For a repeatable small-context check, the guide’s starting context is 4,096 tokens. Confirm the actual loaded setting rather than assuming that a download uses the setting you intended. Context includes instructions, input and conversation history; a whole book will not fit simply because the model has more parameters.

Try a small task without personal data

These synthetic examples are prompts you can try, not reported benchmark results from this video.

Extract facts from a short message

Read this invented message. Return the date, time and meeting location in three labelled lines. Do not add information.

Hi Sam, our appointment is on 18 November at 14:30 at 22 Example Street. If you arrive early, wait at reception.

Expected facts: 18 November; 14:30; 22 Example Street. The year is not given.

Compare information you supply

I need a laptop for writing and occasional video editing. My budget is £900. Compare these fictional options using only the information supplied. Explain the trade-off and say what you still need to know.

A: £800, 16 GB RAM, 512 GB storage, integrated graphics, 1.3 kg. B: £900, 16 GB system RAM, 512 GB storage, dedicated GPU with 8 GB VRAM, 2.2 kg.

A useful answer should preserve the separate system RAM and VRAM figures and ask about unspecified details such as battery life and processor. It should not invent benchmark results.

Explain a small bug

This function should return the first item, but it fails on an empty list. Explain why and suggest a version that returns None for an empty list. Do not modify any files.

def first_item(items): return items[0]

A text-only model can propose a change. Reading your project, changing files and running tests require an app or agent with those tools and permissions; parameter count alone does not establish tool reliability.

Record the result

  1. For each model, record the exact tag, context setting, CPU/GPU placement, whether your normal apps remained usable, the time you actually waited, and whether the answer was right. Keep the same input when comparing two models. Start a fresh chat for each run.

Sources

Ollama download (https://ollama.com/download)

Ollama FAQ (https://docs.ollama.com/faq)

This guide follows the project hardware table and recorded narration. Its recommendations are starting points to check on your own computer, not benchmarks of every configuration.

The takeaway

Start with a model that leaves your laptop usable, check its real memory and CPU/GPU placement, and compare useful answers on the same small task. The hardware table supplies starting candidates, not universal limits.

About this article and its sources

This guide follows the project hardware table and recorded narration. Its recommendations are starting points to check on your own computer, not benchmarks of every configuration.

Original YouTube source ↗