MergeKit

MergeKit, from Arcee AI, is the most widely used open-source toolkit for merging pretrained large language models into new models — no training run required. With around 7,400 stars on GitHub and active development in 2026, it is the tool behind a large share of the merged community models published on Hugging Face.

Overview

Model merging combines the weights of several fine-tuned models to blend their strengths: the coding skill of one, the instruction following of another, the language coverage of a third. MergeKit implements the popular merging algorithms from the research literature — including linear averaging, SLERP, TIES, DARE and Model Breadcrumbs — and drives everything from a single YAML configuration file that declares the models, the method and its parameters. It is engineered to be practical: lazy tensor loading and low-memory execution mean many merges can run on a CPU or a modest GPU, and the output is a standard model directory you can upload to Hugging Face or serve with any inference engine. The toolkit is released under the LGPL-3.0 license.

Key Features

• Implements leading merge methods: linear, SLERP, TIES, DARE, Model Breadcrumbs and more
• Declarative YAML configs that make merges reproducible and shareable
• Memory-efficient lazy loading — many merges run on CPU or small GPUs
• Supports Llama-family and many other popular open model architectures
• Outputs standard model folders ready for Hugging Face or local serving
• The de facto standard tool behind community merged models

💳 Pricing

MergeKit is completely free and open source under the LGPL-3.0 license. There is no paid tier; Arcee AI maintains it as a community tool, and the only cost is the compute you run it on.

Who Is It For

MergeKit is for LLM experimenters, open-model builders and researchers who want to create stronger models by combining existing fine-tunes instead of paying for full training runs, and for teams prototyping specialized models quickly.