llama-swap is a model management tool that enables hot-swapping between multiple generative AI models running on local servers compatible with OpenAI and Anthropic APIs.
The tool solves the problem of running multiple large language models on resource-constrained local machines where only one model can be loaded at a time. It works by intercepting API requests and managing which model is currently loaded, automatically swapping models in and out of memory on demand. This allows users to maintain a single API endpoint while seamlessly switching between different models without restarting services or manually managing model loading.
The project suits developers and researchers running local AI workflows who need to experiment with or deploy multiple models but lack the hardware to run them simultaneously. It works with any server implementing OpenAI or Anthropic API compatibility, including llama.cpp and vllm, making it broadly applicable across the local LLM ecosystem. The tool is designed for simplicity: it requires only a single binary and one configuration file to get running, with no external dependencies, making it accessible to users who want straightforward model management without complex infrastructure.
Development activity shows consistent maintenance with passing continuous integration checks. The project maintains an active presence with regular updates to its codebase. The tool has accumulated meaningful adoption, indicating that the implementation is stable enough for production use in local AI workflows. The straightforward architecture and minimal dependency footprint suggest the maintainers prioritize reliability and ease of deployment over feature complexity.
