Definition
Plain language
An open mixture-of-experts model with 30 billion total parameters but only about 3 billion active per token.
As stated in the literature
A 30B/3B-active sparse mixture-of-experts open-weight base model used as the backbone for SU-01's olympiad-math post-training.
Why it matters: Sparse open backbones make ambitious post-training (like olympiad math) tractable for groups without frontier-scale compute.
For example, only ~3B parameters fire per token even though the full model is 30B, keeping inference cheap during long math rollouts.