On Thursday, Windsurf, a startup known for developing AI tools for software engineers, announced the launch of its first family of AI software engineering models, named SWE-1. The new models—SWE-1, SWE-1-lite, and SWE-1-mini—are designed to optimize the entire software engineering process, beyond just coding.
The introduction of Windsurf’s in-house AI models is noteworthy, especially amid reports of OpenAI’s $3 billion deal to acquire Windsurf. This model launch indicates Windsurf’s ambition to expand from merely developing applications to creating the models that drive them.
According to Windsurf, the largest and most capable model, SWE-1, performs competitively with models like Claude 3.5 Sonnet, GPT-4.1, and Gemini 2.5 Pro on internal programming benchmarks. However, it does not meet the performance of frontier models such as Claude 3.7 Sonnet in software engineering tasks.
Windsurf states that the SWE-1-lite and SWE-1-mini models will be accessible to all users on its platform, whether free or paid, while SWE-1 will be exclusive to paid users. Although the pricing for SWE-1 models has not been disclosed, Windsurf claims they are more cost-effective to serve than Claude 3.5 Sonnet.
Windsurf is recognized for its tools that enable software engineers to write and refine code through interactions with an AI chatbot, a process referred to as “vibe coding.” Other notable companies in this space include Cursor and Lovable, with Windsurf traditionally relying on AI models from OpenAI, Anthropic, and Google.
In a video announcement of the SWE models, Windsurf’s Head of Research, Nicholas Moy, highlighted the company’s distinct approach. Moy stated, “Today’s frontier models are optimized for coding, and they’ve made massive strides over the last couple of years. But they’re not enough for us… Coding is not software engineering.”
Windsurf notes in a blog post that while other models excel at writing code, they encounter difficulties in working across multiple surfaces, such as terminals, IDEs, and the internet. SWE-1 was trained using a new data model and a “training recipe that encapsulates incomplete states, long-running tasks, and multiple surfaces.”
The startup describes SWE-1 as its “initial proof of concept,” indicating potential future releases of additional AI models.
