Windsurf, a startup that uses AI to code music, launches its own AI models[19459016]

On Thursday, Windsurf – a startup that creates popular AI tools aimed at software engineers – announced the launch of its latest product, SWE-1, the first family of AI software-engineering models, was launched . The startup claims that it has trained its new AI models, SWE-1, SWE-1 lite, and SWE-1 mini, to be optimized for “the entire software engineering process,” rather than just coding.

For some, the launch of Windsurf’s in-house AI model may be a surprise given that OpenAI’s AI models have reportedly been around for a while. Closed a $3 billion acquisition deal for Windsurf. This model launch indicates that Windsurf is looking to expand beyond developing applications and into developing the models which power them.

Windsurfstates that SWE-1, which is the largest and most powerful AI model in the group, competes with Claude 3.5 Sonnet GPT-4.1 and Gemini 2.5 Pro when it comes to internal benchmarks. SWE-1, however, appears to fall behind frontier AI models such as Claude 3.7 Sonnet on software engineering tasks.

Windsurf has announced that its SWE-1 lite and SWE-1 mini models will be made available to all users, whether they are paid or not. SWE-1, on the other hand, will only be available for paid users. Windsurf has not announced pricing for the SWE-1 models, but claims that it is cheaper to service than Claude 3.5 Sonnet.

Windsurf’s tools allow software engineers to write code and edit it through conversations with AI chatbots, a practice called “vibe coding.” Cursor is the largest startup in this space, along with Lovable. These startups, including Windsurf have relied on AI models developed by OpenAI, Anthropic and Google for years to power their applications. Windsurf Head of Research Nicholas Moy’s comments in a video announcing SWE models highlight Windsurf’s latest efforts to differentiate itself. Moy says that today’s frontier models have made huge strides in the last two years. They are optimized for coding. “But they’re still not enough for us… Coding isn’t software engineering.”

Windsurf writes in a blog that other models, while good at writing code are unable to work across multiple surfaces, as programmers do, such as terminals and IDEs. The startup claims that SWE-1 was developed using a new model of data and a “training formula” that encompasses incomplete states, long-running task, and multiple surfaces.

According to the startup, SWE-1 is its “initial concept proof,” and it may release other AI models in future. Maxwell Zeff, a senior reporter for TechCrunch who specializes in AI and emerging technology, is available at

www.aiobserver.co