How to Run a Free AI Coding Assistant Locally with VS Code, opencode, and LM Studio

2026-09-05T00:00:00+00:00

Every AI coding assistant wants your credit card. And your code.

There’s a third option, and it costs nothing: run the whole thing on your own computer. No subscription. No internet. Your code never leaves your machine.

I set this up on a fairly ordinary PC, and it works. Here’s exactly how.

What you’re building

Three pieces working together:

LM Studio — runs the AI model on your PC. This is the engine.
opencode — the coding assistant that reads and edits your files.
VS Code — where you actually write code.

LM Studio does the thinking. opencode does the work. VS Code is where you sit.

My setup: Windows, RTX 3070 Ti with 8GB VRAM, 16GB RAM. That’s a mid-range gaming PC, not a workstation. If yours is similar, you’re fine.

Step 1: Install LM Studio

Download it from lmstudio.ai and install it like any normal app.

LM Studio lets you download and run open-source AI models directly on your computer. It’s the easiest way into local AI — no command line required.

Step 2: Download the model

Open the search inside LM Studio and look for qwen3 8b.

You’ll see a lot of versions. Check these three things before you download:

Format: GGUF
Quantization: Q4_K_M
Capabilities: “tool use” must be listed

Why those matter, in plain English:

Quantization is compression. Q4_K_M shrinks the model so it fits on a smaller graphics card. You lose a little quality, but you gain the ability to actually run it.

Tool use is non-negotiable. A coding assistant needs to open your files and edit them. A model without tool use can only chat about your code — it can’t touch it. Skip this check and nothing will work later.

Got different hardware? Pick a model that fits it. Bigger models are smarter but hungrier. An 8B model is a comfortable fit for 8GB of VRAM.

Once the download finishes, your model shows up under My Models.

Step 3: Load the model

Go to the Developer tab and select your model.

Turn on “Manually choose model load parameters”, then click the small arrow next to the model name to open the settings.

Now find context size and set it to 16000.

Context size is how much text the model can hold in its head at once — your question plus its answer plus whatever code it’s looking at. Bigger context means it understands more of your project. It also eats more VRAM.

16000 is a good number for 8GB. If you have less, go lower. If the model refuses to load, lower it again and try once more.

Step 4: Turn on the server

Flip the server status to Running.

Your model is now live at http://127.0.0.1:1234.

That address is your own machine talking to itself. Nothing is going out to the internet — which is the entire point.

Step 5: Install VS Code and opencode

Install VS Code if you don’t have it.

Then install opencode. On Windows in command prompt, the simplest route is npm:

npm install -g opencode-ai

Open a terminal inside VS Code (Terminal → New Terminal) and type:

opencode

opencode starts up right there in the terminal panel.

Step 6: Point opencode at your local model

Here’s the part that trips people up. opencode has no idea your model exists yet. You have to tell it, using a config file.

Create the config file — here’s mine

save it to:

C:\Users\YOUR_USERNAME\.config\opencode

Swap YOUR_USERNAME for your actual Windows username. If that folder doesn’t exist, create it.

Step 7: Pick your model and test it

Restart VS Code, then start opencode again.

Type /models and select qwen/qwen3-8b from the list.

Now ask it to do something real. Give it a file to fix.

Want proof it’s actually running locally? Switch over to LM Studio and check the logs. You’ll see tokens streaming as the assistant types.

That’s your own GPU doing the work.

What to expect

Let’s be straight about this: a local 8B model is not going to match Claude or GPT on a hard architectural problem. It’s smaller, and smaller means less capable.

But for the everyday stuff — writing boilerplate, explaining unfamiliar code, catching bugs, renaming things across files — it holds its own. And it’s fast, because there’s no network round trip.

The best part is what it costs: nothing, forever. No token limits. No monthly bill. Works on a plane.

If you have more VRAM than I do, try a larger model. Same steps, better results.

How to split image dataset into train, validation and test set?

2023-04-22T00:00:00+00:00

Splitting image data into train, validation, and test sets is a crucial step in machine learning model development. It helps to prevent over-fitting, evaluate model performance, and ensure that the model generalizes well to new, unseen data. It’s useful to have code that can quickly separate image data into training, validation, and testing datasets. The folder structure would look something like this:

Obtain the image data Retrieve all the images located in the designated folder.

Split the data Randomly separate the image data into three sets: 70% for training, 15% for validation, and 15% for testing.

Copy the images into their respective folders Create separate folders for each set, including the training, evaluation, and testing sets. Copy the images into their respective folders based on the random split.

Full code

“ Why did the machine learning model go to therapy? Because it couldn’t decide between the train, validation, and test sets — it just kept overthinking! :) ”

UNDERFITTED