Windsurf, a startup known for building AI tools for software engineers, has launched its first family of in-house models called SWE-1, SWE-1-lite, and SWE-1-mini. These models are designed to support the full software engineering process, going beyond basic code generation.
This move comes as a surprise to some, especially following reports that OpenAI is acquiring Windsurf in a multibillion-dollar deal. By releasing its own models, Windsurf is signaling its intention to expand beyond building applications to also creating the core AI models behind them.
SWE-1, the most capable model, reportedly performs on par with models like Claude 3 point 5 Sonnet, GPT 4 point 1, and Gemini 2 point 5 Pro in internal programming evaluations. However, it does not yet match the newest frontier models such as Claude 3 point 7 Sonnet in some tasks.
The company stated that SWE-1-lite and SWE-1-mini will be available to both free and paid users, while SWE-1 will be limited to paying customers. Windsurf has not yet shared pricing but claims that SWE-1 is more cost-efficient to run than comparable models.
According to Windsurf, current AI models focus too much on writing code and fall short when dealing with real-world software engineering challenges like switching between terminals, IDEs, and web resources. Windsurf says SWE-1 was trained using a new approach to better handle incomplete tasks, long processes, and multiple work environments.