larionovavi-stack/genesis2-cascade-moe is a CPU-only cascade MoE project that turns system administrator (sysadmin) and network questions entered in natural language into shell commands. Cascade MoE (mixture of experts) refers to a structure that chains several specialized modules together according to the input to produce an answer. The repository description states that it uses neither a GPU, nor the cloud, nor external model weights. The commit a1221ef at 2026-11-07 01:35 UTC has the message 'chore: demo restarted (keeps scheduled workflow alive)', and the only modified file is .github/last-run.txt.
The actual change in this commit is just one line in that file. '2026-11-06 17:48 UTC' was changed to '2026-11-07 01:35 UTC', and the commit stats show 1 insertion and 1 deletion. In other words, this is a trace of the demo being restarted so that the scheduled workflow would keep running, not a change to the model architecture or source code. As the file name (last-run.txt) suggests, this project is operated as a setup that automatically runs the demo on a periodic basis.
The project itself focuses on moving system administration tasks from natural language into shell commands. The repository description states that it produced commands that worked on 80% of held-out Russian queries, while English was 38%. It is a Russian-first (Russian-first) design, and the benchmark used for evaluation is included alongside the source. Patents are marked as being at the patent pending stage (patent pending).
To be clear about the limitations, this commit amounts to no more than updating a single line of the execution record and contains neither source changes nor performance improvements. The performance figures likewise rely on the repository's own benchmark, which the repository also provides, and on the repository description, so they cannot be treated as externally verified results. This record contains no license and no actual adoption case. Therefore it is a case worth checking first for someone looking for a shell command generation tool that runs on CPU alone, and the performance in other languages such as English and the license terms are stages that require separate confirmation.