SERA – Fast, adaptable open coding agents for any repository

Summary: SERA is a family of open-source coding models (8B, 14B, 32B) trained using a novel efficient method on "soft-verified" data, significantly reducing training costs. It supports easy adaptation to private codebases and provides open weights, data, and recipes for custom development.

What it does

SERA models learn from partially correct code patches to improve training efficiency and enable specialization to specific codebases or internal frameworks.

Who it's for

Developers and teams seeking customizable coding agents for private or specialized repositories.

Why it matters

It lowers synthetic data costs and enables efficient training of coding models tailored to unique codebases.