The Allen Institute for AI launched SERA, open-source AI coding agents that outperform larger models and cost just $400 to train.

The Allen Institute for AI has released SERA, a family of open-source AI coding agents that outperform many larger closed models on code benchmarks while costing as little as $400 to train. SERA-32B and SERA-8B use a novel soft-verified generation method to create high-quality training data cheaply, enabling strong performance on real-world codebases without expensive infrastructure. The agents can be fine-tuned on private code, integrate with tools like Claude Code, and deploy with a single command, making advanced AI coding tools accessible to small teams and individuals.