justvugg/colibri is a pure-C inference engine that runs mixture-of-experts (MoE) models without a graphics device — an approach that splits a huge model into several smaller submodels and computes only the ones that are needed. Even with MoE models in the hundreds of billions of parameters, it loads only the required experts into RAM and reads the rest from disk. The latest release published on GitHub is v2.0.0.
The README sets the starting point for running at 8GB of memory and 22GB of free disk space. The installer picks a model that matches the machine's specifications and carries on with the download, and it provides a browser dashboard along with OpenAI-compatible and Anthropic-compatible endpoints. The README states that it offers ten engines for language models plus engines for image generation and image judging; there are 13 supported engines in total, including GLM-5.2/5.3, Kimi K3, and DeepSeek V4 Flash. The license is Apache-2.0, the repository was created on 2026-07-01, and it has 43013 GitHub stars.
That said, the record contains no changelog for v2.0.0, and no benchmark figures measuring performance remain. Among the evaluation items, only the momentum score is filled in, at 10.0, and the rest carry no scores. The 8GB and 22GB figures are merely the starting conditions the documentation presents, and the record does not reveal what level of token throughput or output quality to expect. So for now it is better to treat this as a project whose execution path and requirements are confirmed, with performance evaluation deferred until actual measurements exist.