DeepSWE

DeepSWE is a computer program that repairs other programs. You give it an error report from a real software project, such as “the date is displayed incorrectly.” The program then independently searches through the project’s files, locates the faulty spot, and writes a correction. The name combines “deep” for the underlying learning technique with “SWE” for Software Engineering, meaning the work of programmers. DeepSWE was released in 2025 by the company Together AI together with a research group from the University of Berkeley. Unlike many well-known AI systems, it is freely available: anyone may download it, examine it, and modify it.

Why an open model for coding work caused a stir

Until then, the rule was: anyone wanting an AI that independently handles programming work had to pay large American providers. Their systems are trade secrets. No one from outside can verify how they were trained. DeepSWE showed that a comparatively small, open model can push into the same performance class.

This is measured using a test called SWE-Bench Verified. It contains hundreds of real bugs from public software projects whose correct solutions are known. DeepSWE solved around 59 percent of these on its own. At the time of its release, that was the best result among freely available models of this size. For comparison: as recently as 2023, the best systems scored under five percent.

This is economically relevant because software maintenance is expensive. A large share of developers' working time goes not into building new things but into fixing old bugs. If a model takes over even part of that, it shifts cost calculations across entire industries. And because DeepSWE is open, companies can run it on their own machines without sending their source code to a provider.

Learning through reward instead of copying

Most language models learn by reading vast amounts of text and predicting which word comes next. DeepSWE does start out with such a pretrained model, but afterward it took a different path. It was further trained using reinforcement learning. This means: the model tries out a solution, and the outcome is evaluated. Good outcomes are rewarded, bad ones are not.

What matters is where this evaluation comes from. It does not come from humans, but from the software project’s automated tests. If the test passes after the change, the fix was correct. If it fails, the fix was wrong. This feedback is unambiguous and can be repeated millions of times without anyone having to watch. That is the reason programming is particularly well suited to this type of training.

A comparison helps: a student can learn mathematics by memorizing other people’s solution methods. Or by calculating things themselves and afterward checking from the result whether it is correct. The second method is slower, but it produces genuine understanding of the intermediate steps. DeepSWE works according to the second principle. In the process, behaviors emerged that no one programmed: the model checks its own changes afterward and discards approaches that lead into a dead end.

DeepSWE in the developer’s toolbox

Direct contact with DeepSWE is rare, since it is not a chat program for everyday use. It runs inside so-called coding agents. These are programs that give the model access to files, search functions, and a command line. The model then decides step by step which tool to use next.

In the news, the name usually comes up in connection with two debates. The first revolves around open versus closed AI models. The second concerns the question of how much programming professions are changing. DeepSWE often serves as evidence that the gap between freely available and commercial technology is narrowing.

A common misconception is that such a model could write complete programs from scratch. It cannot do this reliably. Its strength lies in clearly defined tasks within existing projects for which a test exists. If this objective benchmark is missing, the success rate drops considerably. Even for the tasks it solves, a human ultimately reviews the change before it is incorporated into real software.