The Allen Institute for AI (Ai2), a nonprofit research group founded by the late Microsoft co-founder Paul Allen, is releasing open coding models and a fully open framework for building and training one’s own coding agents.
Most powerful coding agents today are closed, expensive to train and poorly suited to private or internal code, limiting adoption beyond large, well-funded labs. Ai2’s open agents could lower the cost and barriers to deploying AI software assistants by developers and small organizations.
That’s because closed models haven’t seen your internal code, so they don’t know the custom data pipelines, internal APIs and conventions. Training them on private data helps but generating synthetic data from private codebases that actually works is “difficult and expensive,” the nonprofit said.
“Our method makes it easy,” Ai2 said in a blog post.
The first release in its coding agent family is SERA, or Soft-verified Efficient Repository Agents. It solves about 55% of SWE-Bench Verified problems while needing only eight GPU days or fewer to train on two Nvidia H100s or RTX 6000s. SERA delivers fast inference and is compatible with Claude Code, according to the nonprofit.
The system relies on synthetic training data that mirrors developer workflows rather than perfectly verified code, allowing smaller models to match or exceed much larger ‘teacher’ models when specialized to a given repository. All models, code and training recipes are released openly, enabling developers and researchers to adapt coding agents to internal software stacks at low cost, Ai2 said.