Hummingbird
A zero-dependency C runtime for running massive open-weight Mixture-of-Experts language models on consumer hardware by streaming expert weights across SSD, RAM and VRAM as one memory hierarchy.
Hummingbird is a runtime, written in portable C17 with no third-party dependencies, for running large open-source language models, including huge Mixture-of-Experts models such as GPT-OSS 120B, GLM and DeepSeek, on hardware that does not have enough memory to hold them outright. Instead of requiring the whole model to fit in RAM, it keeps always-needed weights resident, streams the much larger pool of expert weights from disk on demand, and caches recently used experts, unifying SSD, RAM and VRAM into one managed memory hierarchy.
It is inspired by an existing specialized engine called Colibri, but aims to be a generalized, modular framework that can run any model architecture through an adapter interface rather than one built for a single specialized case. It installs through a one-line installer script on Linux and macOS or by building from source with CMake and a C17 compiler, and its command-line interface supports one-shot text generation as well as an interactive chat mode.