The Allen Institute for AI (Ai2) has launched SERA, a groundbreaking open-source family of coding agent models designed to be both affordable and efficient. Achieving state-of-the-art results on the SWE-Bench Verified dataset, SERA showcases its power by solving 54.2% of the benchmark problems. This performance is comparable to that of much larger models, despite SERA being a 32 billion parameter model.
One of the most impressive features of SERA is its cost-effectiveness. Thanks to methodologies such as soft-verified generation and leveraging synthetic data, these models can be reproduced for approximately $400, and customized to fit private codebases for around $1,300. This is a significant reduction in cost, making advanced coding agents more accessible to small teams and independent developers.
SERA’s performance is further enhanced by a thorough taxonomy of 51 bug patterns. This approach helps the model to not only understand but also predict and rectify issues effectively in various code repositories. The models run efficiently on Nvidia hardware, capable of processing up to 8,600 tokens per second on Blackwell, demonstrating both speed and capability.
The comprehensive release includes detailed training recipes, models, and data, which ensures ease of reproducibility. Furthermore, Ai2’s efforts have enabled custom coding agents to be fine-tuned on private codebases using just two Nvidia H100 GPUs, paving the way for innovative applications in software development.
Overall, SERA represents a significant leap forward in making sophisticated coding agents more accessible and customizable, while also being economically feasible for a wider audience.

Collections

The Allen Institute for AI released SERA, an open-source family of 32 billion parameter coding agent models that achieves 54.2% success on SWE-Bench Verified. The models can be reproduced for $400 and customized for private codebases for $1,300, using techniques like soft-verified generation and synthetic data. SERA leverages a taxonomy of 51 bug patterns and runs efficiently on Nvidia hardware at up to 8,600 tokens per second. Custom agents can be fine-tuned on private codebases using just two H100 GPUs, making advanced coding agents accessible to smaller teams.

SERA: Efficient, Accessible Open-Source Coding Agents for Custom Codebases