Models aren’t good enough to work on their own, especially with all the nuance in bigger projects. Even while avoiding usage caps on Antigravity, you still need to give it your whole project for it to work. Basically, it feels like the only way to get a reliable AI agent for programming is through a paid extension. However, instead of paying someone else, you should run your own AI.

Small models are a better idea for laptop users

If you have the equipment, go for it

Credit: Jorge Aguilar / How To Geek

Most developers think running AIs needs a large server farm or an overpriced graphics card. When I talk to people about running local models, I usually hear complaints about not having a supercomputer to handle the heavy processing load. If you try to load a giant seventy-billion-parameter model on a five-year-old low-end laptop, your machine won’t work and give you words at a very slow pace. Standard processors have much lower memory bandwidth than dedicated graphics cards. This leads to big problems when moving data around.

Since the processor struggles to read the large files fast enough, generating a single line of code can take minutes. You have to run the model based on what you are working with. There’s a sweet spot where a local model is actually helpful instead of being a frustrating memory hog. For example, Qwen 2.5 Coder has models with 1.5 billion or 3 billion parameters. These sound like a lot, but they’re actually compact versions that need surprisingly little system memory.

A 1.5 billion parameter model compressed with quantization techniques takes up only about two gigabytes of memory, letting you run it smoothly alongside your code editor. You get the benefits of an intelligent pair programmer without needing an expensive dedicated graphics processor. When you use a model of this size, it can easily fit everything into your system memory.

Best of all, it’s only on your CPU. That was my biggest hurdle. I’ve got VRAM and a graphics card, but they aren’t competitive by any stretch of the imagination. By adding a small open-weights model to your local environment, you can have an assistant that doesn’t need an internet connection or a large hardware budget.

Setting up your local AI with GPT4All

This gives you more freedom than LM Studio

While I started with LM Studio, I’d recommend GPT4All as your localized chatbot, because I’ve noticed it is actually smoother. Just go to the GPT4All website, download the installer for your operating system, and run it. It doesn’t ask for any complex terminal commands or Python dependencies during setup. You’ll want to make sure your computer has a good amount of system memory.

You can do this with a spare server or by other means if you have them. It’s possible to run the smallest models with 8 gigabytes, but don’t try to use anything else at the same time. Once installed, you can go to the search bar right on the main screen to look up any model available from the Community Models Explorer tab.

For a CPU-only setup, finding the right model size and format is critical. You can’t load a massive model and expect it to run well without a graphics card. A good choice for coding on standard hardware is a smaller version of the Qwen2.5-Coder family, like the 1.5B or 7B instruction models.

When you search for Qwen2.5-Coder inside GPT4All, you’ll see a list of different quantizations. Quantization compresses the model weights so they’ll fit into your system memory more easily. The q4_0 quantization is typically the best balance of speed and coding capability, which means the model drops to a fraction of its original size while keeping the quality. You’ll just click the download button next to the file, wait for the download to finish, and the model will be ready to load.

However, before chatting, you should adjust a few basic parameters to get the model running well on your processor. Click the Models icon to open the local models view and select your downloaded Qwen model. On the right side of the interface, you’ll see hardware settings. Since you’re not using a graphics card, go to the Device menu and select CPU.

This tells the app to run all the computation layers strictly on your central processor. Next, you’ll need to adjust the context window, which is the model’s short-term memory that holds your code and conversation history. A context length of around 4096 tokens works well for CPU processing. If you set the context too high, the application uses up all your system memory and becomes painfully slow.

Keep everything local

I like to own my own technology

Since the model weights are stored on your local disk, it doesn’t need an internet connection to run. You get real-time code suggestions and chat features right in your development environment without paying monthly fees or sending your private code to an external cloud server.

I also make sure to use a separate older tower as a server. Plus, I use my router as a way to transfer the data. This was something I had lying around, and I have far too many Ethernet cables to count. While not everyone can use this technology today, most computers are built so that many people can. It didn’t take a supercomputer to program what I needed.

Debugging was what sold me on this idea, because I hate scanning text to find a small error. I would use regular AIs, but I don’t like having every AI know my projects and what I am doing. When a script fails or throws an unexpected error, you can paste the stack trace directly into the local chat window, and it stays on your computer.

The assistant can find syntax issues and point out logical flaws in a specific file. It explains the root cause of the bug and suggests code fixes you can add with one click. Just remember that if you want more power, it’s up to you to add whatever you need yourself.

We don’t need to spend any more money

With the rising cost of chips, new computers have gotten much more expensive. Fortunately, you don’t need a new model or to upgrade your setup. You just need your CPU and maybe some extra equipment if you want to get even more serious. Don’t pay another dollar to a cloud operator until you’ve tried it yourself.

Brand

UGREEN

CPU

Intel 12th Gen N-Series

Memory

8GB (Upgradeable to 16GB)

Drive Bays

2 x 22TB