New to ArcFlare? Here's everything you need to know

ArcFlare lets you run AI models on your own computer. Chat with them, let one write and fix code for you, turn a photo into a 3D model, or turn text into speech. Nothing is sent to a company's servers, there is no account, and it is free.

This guide assumes you have never done any of this before. Ten minutes, start to finish.

What you need

• A graphics card (GPU) with some memory (VRAM). This is what runs the models. 8 GB is a good start; 16-24 GB runs the really useful ones. Apple Silicon Macs count too: their memory is shared with the GPU.
• Nothing else to set up by hand. The installer brings its own Node.js if you don't have it, and one command fetches llama.cpp, the engine that runs language models, in the right build for your GPU.

No GPU? Small models still run on the CPU, just slowly.

1. Install it

On Windows, in PowerShell:

On macOS or Linux:

Then let it check your setup:

Green ticks are good. If it says llama.cpp isn't installed, get it in one step:

(Already have llama.cpp somewhere? arcflare set-engine <path-to-llama-server> points ArcFlare at it.)

2. Pick a model that fits

This is the step people get wrong. A model has to fit in your GPU's memory, and bigger is not automatically better. Open the shop:

It knows how much memory your GPU has free and marks every model fits, tight or too big. Pick "Fits my GPU" and start with something near the top of the list. Good first choices:

• 8 GB card: qwen3 (8B) or seed-coder for code
• 16 GB card: gpt-oss (20B) or devstral2-small for code
• 24 GB card: qwen3-coder or glm4.7-flash, both fast and good at code

Choose one, pick Download it, and wait. Models are a few GB each.

You can also look on the website: Models spots your GPU and recommends the best model for each task, and every model page has a will it fit? box.

3. Talk to it

Just run:

A menu asks three things: which program to use, which model, and how much "context" (how much it can remember at once). Pick ArcFlare chat to start. Type, press Enter, and the model answers. Type /bye to leave.

4. Let it write code

The agent can read your files, edit them and run commands, in the folder you start it from:

Ask it something like "this test is failing, fix it". Before it runs a command it asks you Allow this?. Say no to anything you don't understand. Already use OpenCode or Codex? arcflare use opencode points it at your local model instead.

5. Beyond chat

Turn a photo into a 3D model (a .glb file Blender opens):

Turn text into speech (a .wav file):

On an AMD card under Windows, read the 3D guide first. It needs one extra step.

6. Use it from your phone

Inside chat or the agent, type:

You get a link. Open it on your phone and you can carry on the conversation from anywhere. The link contains a secret key: treat it like a password.

The words you will see

• Model: the AI itself, a big file of numbers. Each one is good at different things.
• GGUF: the file format ArcFlare runs language models from.
• B (as in 8B): billions of parameters. Roughly, how big and capable a model is.
• VRAM: your graphics card's memory. Models must fit in it.
• Quant (Q4, Q8): a squeezed copy of a model. Q4 is smaller and slightly worse; Q8 is bigger and closer to the original.
• Context: how much text the model can keep in mind at once, counted in tokens (about ¾ of a word each).
• MoE (30B-A3B): "mixture of experts". A big model that only uses a small part (here 3B) per word, so it runs much faster than its size suggests.
• Agent: a model that can use tools (read files, run commands), not just talk.

Keeping it up to date, and removing it

When there's a new version, ArcFlare tells you in one line. Type /update, or run arcflare update.

Want it gone? This shows everything it would remove, sizes included, and asks before deleting anything:

Your downloaded models are kept unless you add --models.

Where next

• Docs: every command, explained
• Model library: what's available and what fits
• Troubleshooting: the problems people actually hit
• Stuck? Open an issue on GitHub with the output of arcflare doctor.