justvugg/colibri v2.1.0 is a release that expands GGUF execution, CUDA and Vulkan paths, multi-model support, and the chat screen. The changes are recorded in the GitHub release notes published on 2026-10-31. colibri is a pure C engine that runs MoE models on hardware the user already owns; the repository describes it as dependency-free, streaming expert weights from disk.
The goal of this project is to run large MoE models on hardware that already has them. The pure C implementation, the dependency-free setup, and the approach of reading experts from disk are the core of the design the repository puts forward. The release notes state that this v2.1.0 expands not only GGUF execution but also the CUDA and Vulkan paths, multiple models, and the chat screen. The provided release summary does not describe the specific operating conditions or the scope of support for each path.
The figures on long-context handling and inference speed are measurements published by the release author. From the public summary alone, the measurement conditions and the comparison baselines cannot be known. It is therefore difficult to read them as independently verified results, or as performance guarantees that apply to other hardware. The change confirmed in this release is the expansion of execution paths, models, and the chat screen, and the published measurements alone are not enough to conclude a general performance improvement.