AI Blocks › Build and ship software
Model merger
Give it two or more tuned models. Get one model that blends their skills, with no new training.
Model merger is a free AI skill for the job "Build and ship software". Give the instructions to an AI assistant that supports the required files and tools, then review its output.
Free · No account · MIT licence · Made by Orchestra Research · Level: advanced
Made by Orchestra Research. See the original on GitHub; we copied it as it is, with its licence.
What it does
One model that blends the strengths of two or more tuned models, built without any new training. Merge a math model with a coding model and a chat model, and test the result in minutes. The skill uses mergekit and a short config file. You choose a method: a plain weighted average, SLERP for two models, or TIES and DARE for several, which trim clashing changes first. You set the weights, run one command, and test the new model on a few benchmarks against its parents. It runs on a CPU, though a GPU is faster. The models must share the same base design.
You give it
- Two or more models built on the same base
- The skills you want to keep
- Weights or a method preference (optional)
- A test set to score the result
You get
- A merge config file
- The command to run the merge
- A merged model folder
- A comparison against the parent models
- Advice on which method to use
Use it when
- You have a model good at code and another good at chat and want both.
- You cannot afford the GPU time to train from scratch.
- You want to try five blends in an afternoon.
- You need to check that a merge did not erase a skill.
How to use it
- Add it to your AI once: download it and add the folder to Claude, ChatGPT or Codex as a skill.
- Ask for the job in your own words. Your AI opens the skill when it fits.
- It works best when your AI can run Python code.