Hey there!
Summary of this email:
- When to use local models for programming (and course on Codely Pro Standard!)
- 🗞️ Cursor's new model is based on Kimi K2.5 and we don't care
- Agentic Programming course on Codely Pro
- Tech joke of the week
🏠 When to use local models for programming
There are tons of local models out there, but few of them are focused on programming.
The most important thing is that they handle tool calling well. Currently, the best ones are:
- 🥇 GLM-5
- 🥈 Kimi K2.5
- 🥉 Qwen 3.5
Another important requirement: knowing whether you can actually run them on your machine.
We recommend using llmfit for that — it analyzes your machine to see which models are compatible.
That said, unless you run the full version of these models, they won't come close to the quality of an Opus 4.6, GPT-5.4, or a Composer 2.
Using local models for programming makes sense in a few cases:
- You're dealing with very sensitive data that can't leave your machines.
- You want to tinker and learn a bit.
- You can't afford frontier models.
- And not much else.
Where local models do make sense is for adding features to your applications.
That's why we created the Local AI: Privacy and scalability course, to explore how to make the most of them in use cases where it actually makes sense.
Course now available on Codely Pro Standard. 🙌
🗞️ Cursor's new model is based on Kimi K2.5 and we don't care
Last week Cursor released its new programming model: Composer 2.
It became popular right away because its benchmarks showed it outperforming Opus 4.6. But then it went viral because they stirred up a bit of controversy: it's not a model built from scratch by them, as their communications might have suggested, but rather it's based on a Chinese Open Source model.
The first rumors that it was based on Kimi K2.5 reached us while we were live on Café con Codely.
Our first impression was: if this is confirmed, it's going to be hilarious, because according to Anthropic, Kimi K2.5 had been built by distilling Opus 4.6.
So Cursor has built a model better than Opus 4.6, because indirectly, they based it on Opus itself.
And after all the noise, Cursor ended up confirming that yes, it's based on Kimi K2.5.
This overshadowed what actually matters: Cursor has released a very powerful model for a fraction of Opus's price. In our experience, it's fairly close to Opus, but doesn't quite match it.
We've been testing it this week and we loved it. Having competition and seeing Cursor step up after all the noise around their price hikes is a great thing.
On top of that, it coincides with what seems like Anthropic tightening Claude Code subscription usage, with users hitting limits sooner.
Tomorrow we'll be live at 9 AM CET covering all this news and more. Join us on YouTube or Twitch!
🎓 Agentic Programming course on Codely Pro
Last week we launched the Agentic programming with AI: Practical fundamentals course after more than 11 editions of our AI for Programming Workshop.
It's a course we've put a lot of care and time into. Everything we share is how we use agents as of today. And it's been very well received.
We want to take this chance to thank everyone who's been responding to the survey about the course. It helps us a lot to take it to the next level.
Next week we'll publish the second part, where we focus on Skills, MCPs, and the way we work. 😊
And since you've made it to this part of the newsletter, here's the joke of the week, which I know you were waiting for:
Why do programmers use mechanical keyboards? So their code is strongly typed! 😂 😂 😂
Cheers!