SERA is a family of open coding models (8B, 14B, 32B) trained with a new efficient method. SERA learns from "soft-verified" data, drastically reducing training costs. Easily adaptable to private codebases. Open weights, data & recipes.