Its own window and taskbar entry. No browser, no address bar.
The Python backend is compiled in. There is no runtime for you to install and no version to match. WebView2 is fetched automatically if your Windows does not already have it.
forge works from a terminal. The
uninstaller takes the entry back out.
Detects your card and downloads the
matching inference engine: about 35 MB for Vulkan, ~500 MB for CUDA. A resumable
download with a progress bar, rather than 500 MB inside the installer. Then you
type a model name — qwen3:4b — and it works out which repository
publishes it and which quantisation fits your hardware.
The Suggestions tab reads your hardware locally and gives reviewable presets for context, KV cache, Flash Attention, GPU layers and MoE. Nothing is applied or downloaded until you choose it.
ForgeWorks searches image and video models, filters by use and resolution, keeps generation history, and lets an agent prepare a request before you press Generate. No Google or FORGE account is required.
Every release publishes a SHA-256 for each file. Hash your copy and compare:
Get-FileHash .\ForgeSetup-x.y.z.exe -Algorithm SHA256
Compare it against the SHA-256
published in the release's manifest.json.
SmartScreen will show “Windows protected your PC”. Click More info → Run anyway. Checking the hash above is the real verification.
Normal Windows uninstall. It asks before deleting your data — models can be many gigabytes, and a reinstall should find them where you left them.
The in-app Update and restart action uses the safe in-place path. If Windows shows an older-install page for a manual run, choose Do not uninstall; the installer will replace the app files and keep your data.
FORGE is a commercial product and its source is not public. The installer above is the only supported way to run it. Versions 2.0.0 to 2.1.1 were published under Apache-2.0 and that grant still stands for those releases.