The Colibri project says it can run seven families of open MoE models, from 744 billion to 2.8 trillion parameters, by streaming experts from disk on user hardware.
According to the post, Colibri is a project written in pure C that runs seven families of open MoE models ranging from 744 billion to 2.8 trillion parameters. Its approach streams experts from disk, and the post says the software runs on the user's own hardware.
The project offers a technical approach to evaluate for running very large MoE models, but the figures and behavior described are claims from the post, not independent validation. If you use AI tools to study or adapt this approach, avoid entering personal data, trade secrets, or credentials, and follow your organization's access and retention rules.