I published Mizarra.
It is an experimental control plane for running huge MoE models such as Kimi K2.7 Code on a single DGX Spark.
It came out of my experiments with 1T-class MoE models that do not fit cleanly into 128GB unified memory: observing which experts / weight regions are likely to be needed during inference, and deciding what should stay resident in memory.
Mizarra 0.0.0 is an experimental public preview.
For safety, it is dry-run focused for now. Live host pinning / mlock-based residency apply is not enabled yet.
Also, @antirez, thank you. I’ve been following your large local MoE inference experiments, and they encouraged me to pursue this direction.
Feedback is welcome.
https://t.co/2oKTlwFyRR