What is mergecraft?

MergeCraft is a local web application that gives people a visual way to combine multiple AI language models into one, a process usually called model merging. Normally this requires writing and editing YAML configuration files by hand through a tool called MergeKit, but MergeCraft replaces that with a drag and drop graph where you can arrange model checkpoints, layer ranges, and other pieces on a canvas, similar to how node-based tools work in some game development software. The configuration it produces still works with the underlying MergeKit system. The application runs entirely on your own computer, with the interface only accessible at your local address unless you choose to download something remote yourself, so your model files stay on your machine. It supports several merging techniques with names like Linear, SLERP, Task Arithmetic, and TIES, along with layer stacking to build a Frankenmerge and Mixture of Experts setups where multiple expert models share the work. Before running an actual merge, it checks that the pieces you have arranged are compatible with each other. Beyond building merges, MergeCraft includes a dashboard for watching jobs run in real time, showing memory and processor usage, a built-in chat interface for testing a merged model once it is ready, and a benchmarking section that can run standardized tests such as GSM8K and MMLU-Pro to measure how well a merged model performs. There is also a memory planner that lets you compare different memory usage profiles before committing to a merge, which matters because merging models can use a large amount of RAM, storage space, and time. Getting started involves installing Python 3.11 or newer, then running a provided startup script for Windows, Linux, macOS, or WSL, which sets up the environment and opens the interface in a browser automatically. The project describes itself as a community preview, meaning it works today but is still actively developing, and it currently offers an English and Brazilian Portuguese interface, with several other language translations of the documentation itself.