Windsurf has introduced a new group of AI models called SWE-1, designed to help with the entire software development process not just writing code.

The SWE-1 Model Family

• SWE-1: Powerful model with reasoning skills like Claude 3.5 Sonnet, but more affordable. Free for a limited time to paid users.
• SWE-1-lite: This lighter version with better performance than the previous Cascade Base. Free for everyone.
• SWE-1-mini: A super-fast, small model used for passive help in the Windsurf Tab. Also free for all users.

Why Build SWE-1?

Because coding is only one part of a developer’s job. Real-world software engineering includes testing, using terminals, debugging, understanding user feedback, and more. SWE-1 is made to help with all of that.

What Makes SWE-1 Special?

• It’s trained to think through incomplete work and longer tasks, not just final results.
• It works across different tools like terminals, browsers, and editors.
• It learns from how real developers work inside Windsurf tools (Cascade, Tab, etc.).

How Good Is It?

• In testing, SWE-1 performed nearly as well as the best models from big AI labs and better than most open-source or smaller models.
• In live user tests, it wrote helpful code people actually kept and used.

The Secret Sauce: “Flow Awareness”

Windsurf’s tools track the full flow of your work: edits, commands, errors, and even copy-pastes. SWE-1 understands this flow and works with you, jumping in when needed and stepping back when you take over.

What’s next?

We believe these findings open up a new direction for improving model performance by adjusting how they are trained and sampled. We’re excited to explore further refinements of these techniques and understand their implications at larger scales.

If you want to work on this problem, we are rapidly scaling our ML research and engineering teams. Apply here.