The first family of AI software engineering models, or SWE-1 for short, was introduced on Thursday by Windsurf, a business that creates well-known AI tools for software engineers. According to the startup, it educated the SWE-1, SWE-1-lite, and SWE-1-mini family of AI models to be optimized for the “entire software engineering process,” not simply coding.
Given that OpenAI apparently concluded a $3 billion deal to acquire Windsurf, some people may be surprised by the introduction of Windsurf’s internal AI models. This model rollout, however, indicates that Windsurf is attempting to go beyond creating applications to creating the models that underpin them as well.
Windsurf claims that SWE-1, the biggest and most powerful AI model among the group, competes with Claude 3.5 Sonnet, GPT-4.1, and Gemini 2.5 Pro on benchmarks related to internal programming. On software engineering tasks, SWE-1 seems to lag behind cutting-edge AI models like Claude 3.7 Sonnet.
According to Windsurf, all platform users will have free or paid access to the SWE-1-lite and SWE-1-mini versions. SWE-1, meanwhile, will only be accessible to those who have paid for it. Although they did not immediately disclose the cost of their SWE-1 models, Windsurf asserts that they are less expensive to service than the Claude 3.5 Sonnet.
Windsurf is most recognized for its “vibe coding” tools, which let programmers write and modify code by having discussions with an AI chatbot. Other well-known vibe-coding startups are Lovable and Cursor, which is the biggest in the industry. Historically, the majority of these startups—including Windsurf—have powered their apps with AI models from Google, Anthropic, and OpenAI.
Nicholas Moy, Head of Research at Windsurf, highlighted the company’s most recent attempts to set itself apart in a video introducing the SWE models. “The frontier models of today have advanced significantly in the past few years and are optimized for coding,” explains Moy. However, they are insufficient for us. Software engineering is not the same as coding.
In a blog post, Windsurf points out that although other models are adept at creating code, they have trouble navigating between various surfaces, such terminals, IDEs, and the internet, as programmers frequently do. According to the startup, SWE-1 was taught with a “training recipe that encapsulates incomplete states, long-running tasks, and multiple surfaces” and a new data model.
SWE-1 is referred to by the startup as its “initial proof of concept,” implying that further AI models might be released later.