Thursday, Windsurf, a startup that develops popular AI tools for software engineers, announcement The launch of its first family of AI software engineering models, or SWE-1 for short. The startup says it has trained its new family of AI models-SWE-1, SWE-1-Lite and SWE-1-Mini-to optimize for the “whole software engineering process”, not just coding.
The launch of Windsurf’s internal AI models can be a shock for some, since Optaai would have concluded an agreement of $ 3 billion to acquire Winsurf. However, this model launch suggests that Windsurf is trying to develop beyond applications simply in development to also develop the models that feed them.
According to WindsurfSwe-1, the group’s largest and most competent AI model, works competitively with Claude 3.5 Sonnet, GPT-4.1 and Gemini 2.5 Pro on internal programming references. However, SWE-1 seems to fail with border AI models, such as Claude 3.7 SONNET, on software engineering tasks.
Windsurf claims that its SWE-1-Lite and SWE-1-Mini models will be available for all free or paid platform users. Meanwhile, SWE-1 will only be available for paid users. Windsurf did not immediately announce prices for its SWE-1 models, but says it is cheaper to serve than Claude 3.5 Sonnet.
Windsurf is best known for tools that allow software engineers to write and modify the code via conversations with an AI chatbot, a practice known as “room coding”. The other popular startups coding for vibration include Cursor, the largest in space, as well as adorable. Most of these startups, including the board to be formed, traditionally rely on OPENAI, Anthropic and Google IA models to feed their applications.
In a video announcing the SWE models, the comments formulated by the Windsurf research manager, Nicholas Moy, underline the new efforts of Windisurf to differentiate his approach. “Today’s border models are optimized for coding, and they have made massive progress in the past two years,” explains Moy. “But they are not enough for us … Coding is not software engineering.”
Windsurf notes in a blog article that if other models are good for writing code, they find it difficult to work between several areas – as programmers often do – such as terminals, IDE and Internet. The startup indicates that SWE-1 has been formed using a new data model and a “training recipe that sums up incomplete states, long-term tasks and multiple surfaces”.
The startup describes SWE-1 as its “proof of initial concept”, suggesting that it could publish more AI models in the future.
