From the source
It is an open model with a hybrid MoE architecture (30B total, 3B active), offering up to 4x higher throughput and 30% faster task completion for agentic workloads.
It supports a 1M-token context window and can be deployed without manual serving configuration.