Vibe-coding startup Windsurf debuts its AI models, aiming to transform software development with tailored, in-house artificial intelligence tools
Windsurf, a software engineering startup that creates popular AI tools, announced the release of its inaugural family of AI software engineering models, or SWE-1, on Thursday.
The startup claims to have trained its new family of AI models — SWE-1, SWE-1-lite, and SWE-1-mini — to be optimized for the “entire software engineering process” rather than just coding.
Given that OpenAI has purportedly closed a $3 billion deal to acquire Windsurf, the launch of Windsurf’s in-house AI models may surprise some. Nevertheless, the introduction of this model indicates that Windsurf is attempting to broaden its scope beyond the development of applications to include the development of the models that underpin them.
SWE-1, the group’s most competent and largest AI model, is competitive with Claude 3.5 Sonnet, GPT-4.1, and Gemini 2.5 Pro on internal programming benchmarks, according to Windsurf. Nevertheless, SWE-1 appears inadequate compared to frontier AI models, such as Claude 3.7 Sonnet, when it comes to software engineering duties.
According to Windsurf, the SWE-1-lite and SWE-1-mini models will be accessible to all users on its platform, regardless of whether they are free or paid. For the time being, SWE-1 will be exclusively accessible to paid users. Windsurf did not immediately disclose the cost of its SWE-1 models; however, it asserts that serving them is less expensive than servicing the Claude 3.5 Sonnet.
Windsurf is most renowned for its tools that enable software engineers to write and modify code through conversations with an AI chatbot, a practice known as “vibe coding.” Cursor, the largest vibe-coding startup in the industry, and Lovable are also popular. Historically, most of these firms, including Windsurf, have utilized AI models from OpenAI, Anthropic, and Google to power their applications.
In a video that announced the SWE models, Windsurf’s Head of Research, Nicholas Moy, emphasized the company’s most recent endeavors to distinguish its approach. “The frontier models of today are designed for coding, and they have made significant progress in the past few years,” stated Moy. “However, they are insufficient for our needs…” Software engineering is distinct from coding.
In a blog post, Windsurf observes that other models are proficient in code writing. Still, they encounter difficulty when interacting with multiple surfaces, such as terminals, IDEs, and the internet, as programmers frequently do. According to the startup, SWE-1 was trained using a “training recipe that encapsulates incomplete states, long-running tasks, and multiple surfaces” and a new data model.
The startup refers to SWE-1 as its “initial proof of concept,” which implies that it may introduce additional AI models in the future.
