I just submitted the Mistral hackathon.
A few hours ago. I built something, shipped something, and now I have thoughts -- not about what I built, but about the tool I used to build it.
This is about Mistral. And I will try to be fair, even though my feelings are mixed.
I Want European AI to Win
Let me be transparent about where I am coming from.
I am Polish. I live in the EU. When it comes to AI, I genuinely want something that is ours. Not out of nationalism -- out of choice. If Mistral or another European model reached the level of Claude or the latest from OpenAI, I would have a real alternative. I would have competitive pressure. I would have options.
So I went in with real goodwill. I wanted to see what Mistral could do on a serious project.
Some Things Worked. Then the Friction Started.
There is a CLI, and there is a vibe coding mode that actually functions. First few prompts? Not bad. I asked, it delivered.
Then I started noticing the friction. Small things. A function that almost worked but needed manual correction. A structure that was 80% right and needed me to fill the rest.
It felt familiar -- in the wrong way.
With Mistral, I was doing what I used to do about a year ago. Providing structure manually. Correcting small weird things. Going back and forth more than I wanted to.
It was faster than writing everything by hand. It was useful. But it felt like 2024.
What Frontier Models Changed
Something genuinely shifted in the last year of AI-assisted coding.
With the current top-tier models, if you invest real time into a session and know what you are building, you can get something production-grade. Not just a prototype with sharp edges.
With Mistral, the old constraints came back. You need to understand its quirks. You need to manage it more carefully.
My Honest Conclusion
The hackathon was fun. Building under pressure, even with friction, is energizing.
But if I am being real: I would not choose Mistral for serious production work right now. Not when the alternative lets me think bigger and execute faster with less overhead.
Mistral has improved since I last used it. The trajectory is right. The foundation is there. It is just not enough yet.
I want to root for European AI. This hackathon made it harder to do uncritically.
If Mistral reads this: please go faster.
Originally published on Digital Thoughts
Top comments (2)
The frustration with Mistral here is real, but I'd push back slightly — a lot of "the model regressed" situations are actually "my prompt was implicitly relying on behavior that changed". Same prompt, different default assumptions baked into the new version.
Testing structured prompts (explicit constraints, output format, examples) makes model comparisons much more reliable. Built a visual tool for this: flompt.dev / github.com/Nyrok/flompt
Some comments may only be visible to logged-in visitors. Sign in to view all comments.