Vibe coding startup Windsurf launches in-house AI models

On Thursday, Windsurf, a startup that develops fashionable AI tools for bundle engineers, announced the motorboat of its archetypal household of AI bundle engineering models, oregon SWE-1 for short. The startup says it trained its caller household of AI models — SWE-1, SWE-1-lite, and SWE-1-mini — to beryllium optimized for the “entire bundle engineering process,” not conscionable coding.

The motorboat of Windsurf’s in-house AI models whitethorn travel arsenic a daze to some, fixed that OpenAI has reportedly closed a $3 cardinal woody to get Windsurf. However, this exemplary motorboat suggests Windsurf is trying to grow beyond conscionable processing applications to besides processing the models that powerfulness them.

According to Windsurf, SWE-1, the largest and astir susceptible AI exemplary of the bunch, performs competitively with Claude 3.5 Sonnet, GPT-4.1, and Gemini 2.5 Pro connected interior programming benchmarks. However, SWE-1 appears to autumn abbreviated of frontier AI models, specified arsenic Claude 3.7 Sonnet, connected bundle engineering tasks.

Windsurf says its SWE-1-lite and SWE-1-mini models volition beryllium disposable for each users connected its platform, escaped oregon paid. Meanwhile, SWE-1 volition lone beryllium disposable to paid users. Windsurf did not instantly denote pricing for its SWE-1 models but claims it’s cheaper to service than Claude 3.5 Sonnet.

Windsurf is best-known for tools that let bundle engineers to constitute and edit codification done conversations with an AI chatbot, a signifier known arsenic “vibe coding.” Other fashionable vibe coding startups see Cursor, the largest successful the space, arsenic good arsenic Lovable. Most of these startups, including Windsurf, person traditionally relied connected AI models from OpenAI, Anthropic, and Google to powerfulness their applications.

In a video announcing the SWE models, comments made by Windsurf’s Head of Research, Nicholas Moy, underscore Windsurf’s newest efforts to differentiate its approach. “Today’s frontier models are optimized for coding, and they’ve made monolithic strides implicit the past mates of years,” says Moy, “But they’re not capable for america […] Coding is not bundle engineering.”

Windsurf notes successful a blog station that portion different models are bully astatine penning code, they conflict to enactment betwixt aggregate surfaces — arsenic programmers often bash — specified arsenic terminals, IDEs, and the internet. The startup says SWE-1 was trained utilizing a caller information exemplary and a “training look that encapsulates incomplete states, long-running tasks, and aggregate surfaces.”

The startup describes SWE-1 arsenic its “initial impervious of concept,” suggesting it whitethorn merchandise much AI models successful the future.