Your laptop can run an AI. The only question is which size fits. 
Find your row — RAM decides, not money:
16GB ▓▓▓▓ qwen3:8b argues back
8GB ▓▓▓ llama3.2:3b writes code
4GB ▓▓ 360m short reply
2GB ▓ 135m instant
Three moves. That is the whole thing.
1 · install — curl -fsSL https://ollama.com/install.sh | sh (mac, linux; on Windows the installer)
2 · run your row — ollama run llama3.2:3b
3 · make it yours — ollama run llama3.2:3b "you are my editor: short lines, no praise"
That is an offline AI on your own machine, nothing leaving the disk. The 3b is the size tag; drop to 135m on a tiny laptop.
Prefer no terminal? Jan is the same thing with a window · llamafile is one file you double-click · koboldcpp is one file built for long chats.
The rest of the ladder, when you want more
The bit nobody says out loud: once answers are free and instant, you stop saving questions for later — and the questions you ask the moment they appear are the ones that actually get answered.
Pick your row, run move 1, then move 2. Ten minutes, and it lives on your machine.
5 « J'aime »
Pick a quantized build that fits your available RAM first, and check the file size before you download it. Keep your prompts short and focused, so the small model has less to juggle at once. Test it on a simple task first, like summarizing a short paragraph, before trusting it with longer work. Store the model file on fast local storage, so loading does not become the slowest part. If an answer looks weak, try a clearer prompt before switching to a larger model.
2 « J'aime »
This row idea is really helpful. I always guessed at model size before, and then wondered why everything crawled. Starting small and checking what actually fits makes a lot more sense.
Try qwen2.5:1.5b it’s pretty good with 8 gb of ram
mine is an old ddr3 16gb. i am not sure if it handles coding
For an old DDR3 16GB machine, coding is doable, just pick a small quantized model first. A 3B model will feel comfortable, and an 8B model in a 4-bit build can work if you keep the editor and the model as the only heavy apps running. Expect slower loading than on DDR4 or DDR5, since the memory bandwidth is lower, but once the model is loaded the typing and editing pace is usually fine. Keep the context window short and work on one file at a time, so the machine is not juggling a whole project at once. Test it on a small task first, like explaining or renaming a short function, and only move to larger files once that feels smooth.
Wow. thanks for your suggestion.
btw, I also have a high end pc but I wanted to know:
is there a local LLM setup that is almost equal or on par with chatGPT, Claude and Gemini when it comes to app coding automation?
Bonne question, et merci de partager votre approche de configuration ! Mon avis est le suivant : les modèles locaux sont déjà très utiles pour l’autocomplétion de code au quotidien, la refactorisation à petite échelle et l’écriture de tests, surtout si vous avez un PC haut de gamme. Choisir un modèle open source dédié au code et le coupler à un plugin d’éditeur local offrira une expérience très fluide. Cependant, pour ce qui est d’égaler complètement les modèles cloud de pointe pour l’automatisation d’applications complexes, il manque encore un petit quelque chose, surtout en matière de collaboration multi-fichiers, de refactorisation importante et de raisonnement sur de longs contextes. Je suggère de faire un test comparatif sur un petit projet : exécutez la même tâche localement et sur le cloud, puis comparez la vitesse, la précision et le nombre de retouches. Cela vous aidera à décider quelles tâches garder en local et lesquelles confier au cloud, pour une approche à la fois simple et efficace. Amusez-vous bien !
1 « J'aime »
!