Gallery

What Forge is

Forge is an open-source, local-first generative-AI workstation for Apple Silicon. It is MIT licensed and written in Python 3.11+, and it runs entirely on the machine in front of you — no cloud round trip, no rented GPU, no API key, no subscription, and nothing uploaded anywhere. Every image on this page was produced on-device.

Why local-first matters

Running a model on your own hardware changes what you are allowed to do with it. Your prompts and your outputs never leave the machine, so there is no third party logging what you made or deciding after the fact that you should not have made it. It also means the thing keeps working when an API is deprecated, a price changes, or a service goes down. The trade is speed and convenience against ownership, and Forge takes the ownership side of that trade deliberately.

What these renders demonstrate

The gallery is not a highlight reel. It exists to show what the pipeline actually produces at a given point in the project, including the runs that did not clear the bar. The Madhubani catalog covers 41 species, the test suite has 128 passing tests, and a multi-seed mflux run came in 60.8% faster than the previous implementation. Those numbers are published whether or not they flatter the project — when a ship threshold was written down before a run and the result came in under it, the verdict goes next to the methodology rather than getting quietly dropped. A benchmark you can only pass is not a benchmark.

Read the source on GitHub, or see the rest of what Kaayko makes.