Until today, using ArcFlare meant opening a terminal. That's fine if you live in one, and a wall if you don't. ArcFlare Desktop is the same engine in a window, for Windows, macOS and Linux.
Download it here. It's free, open source and MIT licensed, and there's no account.
What's in it
- Chat. Pick a model and talk. ArcFlare sizes the context to the VRAM you actually have free, the same way the CLI does.
- Studio: images. stable-diffusion.cpp is built in. If you already use ComfyUI, connect it and run any of your workflows from the app.
- Studio: 3D, edited by asking. Turn a picture into a mesh with Hunyuan3D or TripoSR. Then brush over a part of the model, say the ears, and type "make these bigger" or "smooth this". A local model plans the edit and the app applies it to the mesh. Undo is one key.
- Studio: speech. Qwen3-TTS, Kokoro, Chatterbox and more. Choose a voice, describe the tone, or clone one from a short clip.
- Flows. Chain steps node by node, ComfyUI-style: prompt → image → 3D, or text → script → voice.
- Harnesses. Open OpenCode, Hermes, Codex or the ArcFlare agent pointed at your local model in one click.
One GPU, shared fairly
Chat models and studio jobs both want VRAM. By default the app stops the chat server while a 3D or image job runs, then reloads your model the next time you send a message, so a 16 GB card can do both, just not at the same moment. You can change that in Settings.
It shares everything with the CLI
The desktop app runs the same engine as arcflare. Models you've downloaded, your memory profile, the llama.cpp server and the generator environments are all shared. Set something up in one and it works in the other.
Installing
1. Pick your system on the download page. 2. The builds aren't code-signed yet, so Windows and macOS may warn you. The terminal installers on that page check the download and open without the warning. From the browser, on Windows choose More info → Run anyway in SmartScreen; on macOS open the app once, then System Settings → Privacy & Security → Open Anyway. 3. The first run downloads llama.cpp for your graphics card by itself. Then open the Shop, pick a model that fits your GPU, and start chatting.
Found something broken? Report it. Every report gets read before the next release.