AI2 Releases SERA Open Coding Agents — Customizable for Any Codebase at $400

Summary

The Allen Institute for AI (AI2) released SERA (Soft-verified Efficient Repository Agents), the first family in its Open Coding Agents initiative. SERA models are open-weight coding agents ranging from 8B to 32B parameters, built on Qwen 3 with 32K context length training, designed to be customized for any codebase — including private enterprise repositories — at remarkably low cost. Performance highlights: SERA-32B solves 54.2% of SWE-Bench Verified problems at 64K context (49.5% at 32K), outp