Ai2 releases SERA open-source coding agents that adapt to any codebase for just $400

Ai2 cooks up open-source coding agents with tech equivalent of 'hot plate and frying pan'

The Allen Institute for AI (Ai2) is open-sourcing the recipe and ingredients for advanced coding agents, making them trainable on an organization's own code base at low cost -- a move that could loosen the grip of tech giants on artificial intelligence for software development. Ai2's newly released model, called SERA (Soft-Verified Efficient Repository Agents), is the first in a series of Open Coding Agents from the Seattle-based nonprofit. The move comes as AI agents reshape software engineering. Popular proprietary tools like Microsoft's GitHub Copilot Workspace, Anthropic's Claude Code, and Cursor bring AI directly into coding workflows but often lock users into expensive, closed systems. Illustrating the potential, SERA was built by a small team that included Ai2 research scientist Tim Dettmers and intern Ethan Shen, a PhD student at the University of Washington's Allen School of Computer Science & Engineering who did much of the development work. In a post about the process, Dettmers said most coding agents are built with the equivalent of an industrial kitchen: hundreds of GPUs, complex infrastructure, and large teams. For SERA, he explained, "we had the equivalent of a hot plate and a frying pan: 32 GPUs and five bright-eyed researchers who wanted to cook state-of-the-art coding agents." SERA agents can take GitHub issues or bugs, generate fixes via line-by-line patches, and submit pull requests. After they're fine-tuned on a specific codebase, they develop deep knowledge of internal APIs and software development conventions. Because the underlying model and training code are fully open, teams are able to run it on their own infrastructure without ongoing licensing fees. Ai2 said the strongest version, SERA-32B, handles more than half of tough real-world coding problems from the SWE-Bench test, a popular benchmark. That puts it on par with top closed models, but it's built on fully open technology designed for anyone to download and modify. Software development teams can set it up with a few lines of code, and it works with tools like Claude Code out of the box. Unlike proprietary rivals, SERA can be customized to a company's private code for as little as $1,300 using commodity GPUs on public cloud platforms.

Ai2 launches family of open-source AI developer agents that adapt to any codebase

Ai2 launches family of open-source AI developer agents that adapt to any codebase Artificial intelligence is moving swiftly, changing how developers craft, as code flows ever faster into repositories such as GitHub and machine minds now work alongside human hands. According to the Allen Institute for AI, coding agents suffer from a fundamental problem: Most are closed, expensive to train and difficult to study or adapt to private codebases. To tackle this issue, today the company released Ai2 Open Coding Agents, a collection that makes building and training custom coding agents accessible and simple. The first release within the family, named SERA, for Soft-verified Efficient Repository Agents, can solve more than 55% of SWE-Bench Verified problems, a benchmark surpassing prior open-source models of comparable sizes. Every component of SERA is open, including models, code and integration into Anthropic PBC's Claude Code. It also launched with a single line of code. Users do not need any large language model training experience. Under the hood, SERA arrives in two versions: SERA-32B and SERA-8B. The first is a 32 billion-parameter model that delivers strong SWE-bench Verified performance. It solves about 55% of issues in standard settings, besting most open models such as Qwen3-Coder, and closed models, including Mistral3's Devstral Small 2, on matched inference setups. The second is an 8 billion-parameter model that solves 29.4% of SWE-Bench Verified problems vs. 9.4% on reinforcement learning baselines; for example, models such as SkyRL-Agent-8B-v0 solve at 9.4% using Qwen 3 8B model, whereas SERA-8B reaches the higher rating. Ai2 used specialized models, trained on 8,000 synthetic trajectories per repository, consistently matched and often exceeding the performance of GLM-4.5-Air, a greater than 100 billion-parameter model used as a teacher. One particular result, Ai2 said, that showed particular promise is that the smaller, fully open model could replicate or even exceed the performance of a more capable "teacher" coding agent. Thanks to advantageous specialization and fine-tuning at the 32 billion-parameter level for specific codebases, SERA can surpass some 100 billion general-purpose models at one-third the size. At deployment, this means a smaller memory footprint and much lower compute, resulting in much lower costs, without sacrificing quality. The total cost to reproduce the main results as seen by Ai2 on commodity cloud hardware is around $400, about 100 times cheaper than many existing approaches on the market today. The company explained that the release includes everything developers and researchers need to hit the ground running to reproduce, test and build on SERA: a lightweight deployment based on two lines of code for launch, deployment and inference. There is also a setup script and inference optimization for SERA to work with Claude Code. Ai2 said it intends to use the same recipe to keep improving and scale to larger backbones, but maintained that the current pipeline is already cheap and feasible for anyone to run, customize and iterate on.