Researchers at the Allen Institute for AI (AI2) have unveiled a new family of coding agents called SERA, which stands for Soft Verified Efficient Repository Agents. This innovative approach aims to enhance coding efficiency by utilizing supervised training and synthetic data, making it possible to tackle larger closed systems.
SERA is the first model in AI2’s Open Coding Agents series. The standout model, SERA-32B, is built on the Qwen 3 architecture and functions at the repository level. In tests, SERA-32B achieved a 49.5% resolve rate with a 32K context and improved to 54.2% at 64K context. These results are competitive with other prominent systems, such as Devstral-Small-2 and GLM-4.5 Air, despite SERA being fully open in terms of code, data, and weights.
Currently, the SERA series includes four models: SERA-8B, SERA-8B GA, SERA-32B, and SERA-32B GA. All these models are available on Hugging Face under an Apache 2.0 license.
A key feature of SERA’s training process is Soft Verified Generation (SVG). This method simulates realistic developer workflows to create agent trajectories. It involves two rollouts that produce patches, which are then compared to assess their correctness. This approach allows for soft verification, meaning that even less-than-perfect data can still be useful for training.
The training pipeline for SERA utilized 121 Python repositories from the SWE-smith corpus, generating over 200,000 trajectories. This makes it one of the largest datasets available for coding agents. Specifically, SERA-32B was trained on a subset of 25,000 trajectories, using a standard supervised fine-tuning process.
The research team found that using realistic but sometimes noisy data could still yield effective results. They also highlighted that SERA’s training is significantly more cost-effective than traditional reinforcement learning methods, with estimates showing it is about 26 times cheaper than systems like SkyRL-Agent.
Another interesting aspect of SERA is its ability to specialize in specific repositories. The team tested this on popular projects like Django, SymPy, and Sphinx, generating tens of thousands of trajectories for each. Specialized models showed promising performance, often matching or slightly exceeding the results of their teacher model, GLM-4.5 Air.
In summary, SERA represents a significant step in the development of coding agents. By focusing on supervised learning and utilizing realistic data from actual repositories, AI2 is paving the way for more efficient and effective coding automation. The models are designed to be accessible and practical, making them valuable tools for developers.