I hate vibe coding, but how I feel is irrelevant. The world is a broken hamster wheel, and it spins faster and faster. After OpenAI’s Codex, there’s a new “bug factory” in town, courtesy of the makers of Windsurf, formerly called Codeium.

Windsurf’s SWE-1 family

TechCrunch: Vibe-coding startup Windsurf launches in-house AI models:

On Thursday, Windsurf, a startup that develops popular AI tools for software engineers, announced the launch of its first family of AI software engineering models, or SWE-1 for short. The startup says it trained its new family of AI models — SWE-1, SWE-1-lite, and SWE-1-mini — to be optimized for the “entire software engineering process,” not just coding.

The launch of Windsurf’s in-house AI models may come as a shock to some, given that OpenAI has reportedly closed a $3 billion deal to acquire Windsurf. However, this model launch suggests Windsurf is trying to expand beyond just developing applications to also developing the models that power them.

According to Windsurf, SWE-1, the largest and most capable AI model of the bunch, performs competitively with Claude 3.5 Sonnet, GPT-4.1, and Gemini 2.5 Pro on internal programming benchmarks. However, SWE-1 appears to fall short of frontier AI models, such as Claude 3.7 Sonnet, on software engineering tasks.

But let’s go to the source. May 15: SWE-1: Our First Frontier Models:

Today, we are launching our first family of models, dubbed SWE-1, optimized for the entire software engineering process, not just the task of coding.

This family is currently comprised of three distinct models:

  • SWE-1: Approximately Claude 3.5 Sonnet levels of tool-call reasoning while being cheaper to serve. It will be available to all paid users for a promotional period of 0 credits per user prompt.
  • SWE-1-lite: A smaller model that replaces Cascade Base at better quality. It is available for unlimited use to all users, free or paid.
  • SWE-1-mini: A small, extremely fast model that powers the Windsurf Tab passive experience for all users, free or paid.

Why build SWE-1? Simply put, our goal is to accelerate software development by 99%. Writing code is only a fraction of what you do. A “coding-capable” model won’t cut it.

Some fancy charts:

What you want to know about (if you don’t already) is called Cascade, the smart way of AI-powered software development with the Windsurf Editor (which is actually an IDE) and the various JetBrains IDEs.

For pricing, you can consult Cascade > Plans and Credit Usage, and Plans and Pricing. The Free plan comes with:

  • 1 App Deploy per day
  • 25 prompt credits/month — Equivalent to 100 GPT-4.1 prompts (4 prompts per credit)
  • Unlimited Windsurf Tab
  • Unlimited Previews

Windsurf can still be used for free after your credits are exhausted. When editing code, you’ll still have access to unlimited Tab completions and AI command instructions.

What you need to know is that the full-strength SWE-1 model is free for a limited time! When this offer ends, only SWE-1-lite will remain free to use indefinitely. (SWE-1-mini will remain free in Windsurf Tab as a much smarter IntelliSense.)

Where to start?

If you’re a young, trepaned robot who embraces all the latest technologies, no matter how disruptive they are, you have already started using Codeium long ago. I’m a more conservative person who can’t stand too many “future shocks” in a lifetime and who dreads information overload precisely because I have lots of interests, not just the latest 73,854 JS frameworks and the newest 3,538 AI-powered software development tools. So I installed Codeium, I found it off-putting, and I let it there, only to frown upon its sudden change of name. I find the new name totally retarded!

I also hate visual tutorials on YouTube because most of them are pure crap. But there are some very decent channels (a drop of water in an ocean of garbage), and I have to be fair and admit that the best introduction to Windsurf that I know of can be found on the channel of Volo Builds.

Oh, but before that, while we’re still in text mode, a couple of links to Reddit.

Read this entire comment, not just my excerpt: Windsurf best practices:

I see a lot of folks struggling with Windsurf with issues that I’m not experiencing. Reading through some posts I suspect the main issue is context strategy (or lack thereof). I’m an experienced software engineer (30+ years) and former Googler. I’m currently doing various independent projects.

The approach I take is very documentation driven and not unlike how I would approach projects with a team. I think the “Memento” analogy applies well to the situation. If you’ve seen the movie, it’s about a guy who wakes up each morning with total amnesia. Upon waking and looking around, he sees all of these notes posted around the room that he wrote for himself the night before to bring his new, memory lost self up to speed.

This is how you need to approach coding with LLMs. I probably spend 80% of my time working with my LLM tools documenting what we’ve done and what we need to do. This makes the actual coding work extremely efficient and on point.

Every session begins with an initialization process (finding the reminder notes in “Memento”), and ends with a documentation session (writing the memento notes for the next session).

So this doesn’t need to be “pure vibe coding”!

People can still feel frustration:

I went way deep and got frustrated, windsurf forgets way too often. I think a simpler approach may end up working better. Here are some pieces from the prompt, but one thing was to have at the bottom — “re-read this after every response”, I also had an instruction to put an emoji at the front of the responses so that I could tell if it forgot the context.

Some bugs, maybe?

Using Windsurf a few days, I noticed that even when I put documents for reference in there, it very fast forgets what’s in those documents and constantly needs to be reminded to look it up.

Another funny thing: Several times Windsurf was struggling to fix errors and created new errors instead. I then asked it to act as a lead developer, analyze the root cause and fix it. Then it fixed it right away.

Now, la pièce de résistance From April 17, 2025: Windsurf AI Tutorial for Beginners:

This guy prefers Claude 3.5 Sonnet to Claude 3.7 Sonnet because he finds 3.7 “to overengineer solutions and write too much unnecessary code; so I don’t really use it.” I’ll get back to this later.

How about helping you choose between Cursor AI and Windsurf, ex-Codeium? From March 6, 2025: Cursor AI Tutorial for Beginners [2025 Edition]:

Comment by @etamm1343:

I add a prompt “Think step by step about whether a simpler, more elegant, and robust solution exists that avoids over-engineering while aligning with KISS, YAGNI, DRY, and SOLID principles. Present your solution along with your confidence level (1–10) and a rationale, but do not modify the code yet.” so 3.7 doesn’t overengineer the code.

To be taken into account. Claude is Claude no matter where you’re using it. But even for non-IT tasks, I found Claude 3.7 a bit strange compared to 3.5. I fear the newest Sonnet 4, because for everyday questions it’s by no means smarter.

Of possible interest, from the same guy, on Feb. 6, 2025: AI Coding 101: Ultimate Prompt Guide (37 tips):

Companion document: AI Coding 101: Ultimate Prompt Guide.

A last selected video, his initial complaint about Claude 3.7 Sonnet being worse than Claude 3.5 Sonnet:

“It really likes to be verbose, producing excess code and slowing down iteration.” Maybe it requires different “prompt engineering techniques.” It happened the same with SEO: initially, it was about proper structuring so that search engines would properly identify, classify, and whatnot. Then, it became a way to cheat Google and push your irrelevant scam page up in the result list. But even the creators of legitimate content started to use questionable SEO techniques, because Google was broken. What if Claude is broken?

A last quick comment

Watch carefully the quick Windsurf tutorial. Watch it on YouTube if you want to be able to navigate through the chapters or to read some comments. It’s not longer than it needs to be. It’s not created by a moron. On the contrary, not only does this guy know what he’s doing, but he seems to have quite good taste. In all of his videos that I’ve watched.

But I got terrified. I was already terrified for many years to see the gazillions of stupid JS web apps that make the browser running them use the entire RAM and CPU. I am sickened by the thicket (or should I have said welter?) of frameworks and shit, by the absurdly complex architecture of such apps, and by how the joy of software development went away, except in young, dedicated robots.

This guy is no vibe coder. He seems a professional. And he was quick to fix the junk code issued by the AI. But how about other people?

“Vibe coders” might eventually obtain apps that run, but they will have no clue as to how or even why they are running properly at first sight. I recoil in horror at the thought of gazillions and gazillions of garbage apps created by AI and pushed into production.

The more experienced programmers might still be enthusiastic, like this guy, or like the former Googler with 30+ years of experience. They’ll just fix AI’s mistakes, hallucinations, and regressions, and that would be everything. Atta nice tool!

Some of them, even if they don’t share my repulsion of everything that’s too disruptive, might still feel that the workflow of developing software with AI’s help is unbearably uncomfortable.

At one moment, the AI seems to be extremely helpful, and it creates an entire project in a snap. The next thing you know, it breaks everything just because you asked it to make a small change.

This goes way beyond the regressions made by an honest mistake of a human being. Lots of random, unpredictable crap that you have to work around just because you want to benefit from the moments when the AI is not on LSD.

This is the end of decency on planet Earth.