sqliteai/warp

WARP, formerly called WASTE, is a program written in C that lets you run extremely large AI language models on a regular computer, even when the model is far too big to fit in your computer's memory. It does this by keeping the core parts of the model in RAM and reading the rest of the model's data directly from your hard drive as needed, streaming pieces in only when they are actually required for the current word being generated. The project's flagship example is Kimi K3, a model with 2.78 trillion parameters. Its full set of weights takes up over a terabyte of storage, yet WARP can run the complete, unmodified model on a 64 GB MacBook Pro at roughly half a word generated per second. That speed is slow compared to cloud AI services, but it means a model this size can run at all on hardware a person might already own, without needing a server farm. Kimi K3 is built as a mixture of experts, meaning only a small fraction of its parameters, about 4 percent, are used to answer any single question. WARP takes advantage of this by loading only the specific experts needed for each step, predicting ahead of time which ones will be needed next so it can start reading them from disk before they are required. It compresses the less important parts of the model heavily and keeps the more sensitive parts at higher precision to preserve accuracy. WARP also supports images. You can attach one or more pictures to a prompt and ask the model questions about them, though processing images takes noticeably longer than processing text because each part of an image is treated similarly to a word by the underlying model. Besides Kimi K3, WARP can also convert and run models from the DeepSeek family, including DeepSeek V3, R1, and Kimi K2. To build WARP yourself you need a C compiler, and it runs on macOS, Linux, and Windows, with no other required software dependencies for standard use on a computer's processor.