Windsurf, a startup known for creating AI model for software developers, has launched its first proprietary family of software engineering models, called SWE-1. The suite—which includes SWE-1, SWE-1-lite, and SWE-1-mini—aims to support the entire software engineering workflow, not just code generation.
Windsurf may surprise many with the release of its in-house models, especially amid reports that OpenAI has finalized a $3 billion deal to acquire the company. Nonetheless, the launch shows that Windsurf aims to move beyond application development and build its own foundational AI models.
Windsurf claims that SWE-1, its most advanced AI model, performs on par with models like Claude 3.5 Sonnet, GPT-4.1, and Gemini 2.5 Pro on internal coding benchmarks. However, it doesn’t quite reach the level of top-tier models like Claude 3.7 Sonnet in software engineering capabilities.
The company will make SWE-1-lite and SWE-1-mini accessible to all users, both free and paid. In contrast, it will reserve SWE-1 for paying subscribers. Although Windsurf did not share pricing details at launch, it claims its models are more cost-efficient than Claude 3.5 Sonnet.
Windsurf has gained recognition for enabling engineers to engage in “vibe coding”—a conversational coding style powered by AI chatbots. Other major players in this space include Cursor, the largest in the category, and Lovable. Traditionally, these companies have used third-party AI model from providers like OpenAI, Anthropic, and Google to build their platforms.
In a launch video, Nicholas Moy, Windsurf’s Head of Research, highlighted the rationale behind the new models: “Today’s frontier models are optimized for coding, and they’ve made massive strides over the last couple of years,” Moy explained. “But they’re not enough for us … Coding is not software engineering.”
A blog post from the company elaborates that while existing AI model can write code effectively, they often fall short when it comes to managing tasks across multiple environments—such as terminals, IDEs, and browsers. Windsurf says SWE-1 was trained using a novel data approach and a “training recipe that encapsulates incomplete states, long-running tasks, and multiple surfaces.”
The startup refers to SWE-1 as its “initial proof of concept,” indicating that additional AI models may be in development for future release.
Windsurf’s launch of its own AI models marks a significant step as the company takes greater control over the technology powering its tools. By developing and deploying SWE-1 and its variants, Windsurf aims to deliver a more integrated and efficient software engineering experience for developers, setting the stage for continued innovation in the vibe-coding space.
