On Thursday, Windsurf, a startup that develops common AI instruments for software program engineers, introduced the launch of its first household of AI software program engineering fashions, or SWE-1 for brief. The startup says it skilled its new household of AI fashions — SWE-1, SWE-1-lite, and SWE-1-mini — to be optimized for the “complete software program engineering course of,” not simply coding.
The launch of Windsurf’s in-house AI fashions could come as a shock to some, provided that OpenAI has reportedly closed a $3 billion deal to accumulate Windsurf. Nevertheless, this mannequin launch suggests Windsurf is making an attempt to broaden past simply growing purposes to additionally growing the fashions that energy them.
In keeping with Windsurf, SWE-1, the most important and most succesful AI mannequin of the bunch, performs competitively with Claude 3.5 Sonnet, GPT-4.1, and Gemini 2.5 Professional on inside programming benchmarks. Nevertheless, SWE-1 seems to fall in need of frontier AI fashions, corresponding to Claude 3.7 Sonnet, on software program engineering duties.
Windsurf says its SWE-1-lite and SWE-1-mini fashions might be out there for all customers on its platform, free or paid. In the meantime, SWE-1 will solely be out there to paid customers. Windsurf didn’t instantly announce pricing for its SWE-1 fashions however claims it’s cheaper to serve than Claude 3.5 Sonnet.
Windsurf is greatest recognized for instruments that enable software program engineers to write down and edit code by means of conversations with an AI chatbot, a apply often called “vibe coding.” Different common vibe-coding startups embrace Cursor, the most important within the area, in addition to Lovable. Most of those startups, together with Windsurf, have historically relied on AI fashions from OpenAI, Anthropic, and Google to energy their purposes.
In a video saying the SWE fashions, feedback made by Windsurf’s Head of Analysis, Nicholas Moy, underscore Windsurf’s latest efforts to distinguish its method. “In the present day’s frontier fashions are optimized for coding, and so they’ve made huge strides during the last couple of years,” says Moy. “However they’re not sufficient for us … Coding just isn’t software program engineering.”
Windsurf notes in a weblog put up that whereas different fashions are good at writing code, they battle to work between a number of surfaces — as programmers usually do — corresponding to terminals, IDEs, and the web. The startup says SWE-1 was skilled utilizing a brand new information mannequin and a “coaching recipe that encapsulates incomplete states, long-running duties, and a number of surfaces.”
The startup describes SWE-1 as its “preliminary proof of idea,” suggesting it could launch extra AI fashions sooner or later.
