About once a week I’ll come across a broad question that I think an LLM can answer without hallucination. I’ve recently switched to Mistral’s “vibe” app for that and find it good.
It’s the only moderately sized European AI company so worth your consideration at least.
Mistral is pretty good now. Because they host deepseek v4.1. Also they just recently completed a python=>rust port of their CLI, which is nice. Grandfathered with an annual subscription and you get near pretty nice usage for most coding patterns with very little money.
I’ve been grinding away 24/7 with a deepseek implementer on mistral and glm5.3 reviewer on melious.ai this week and I’ve used about 35 euros of the allotted 255 euro monthly usage. The reviewer needs so few tokens that my melious.ai subscription will soon max out its energy.
Their own models keep being about 1 year behind the competition.
Too bad Mistral doesn’t use Mistral. 😕

Which does not mean that they don’t develop own models.
When you use the coding tab it uses Mistral Large 4 for example. Voice input/TTS is also run on their models.
I really would like to use them but they are still so far behind.

source: https://artificialanalysis.ai/
I hope they can catch up in the future, but for now if you want to buy european and want quality its better to use a chinese model on a european host.
To give them some chance of credit, Mistral claims that they have optimized their model to work much better with their own harness. I don’t know about that, but most benchmarks use a common harness for all AIs so any such optimizations don’t get applied.
Benchmarks don’t tell the whole story though, I’m sure they’re doing their best to be representative and measure something that is in reality pretty random, but I also find they’re pretty synthetic and not totally reliable. As a non-synthetic human I actually find a lot of Mistral’s answers pretty decent and models derived from it tend to score pretty highly on my personal “I find myself using them a lot because I like their output” benchmark. Take that anecdote for what it’s worth (very little) and YMMV, but don’t always completely base your choices on benchmarks either.
That is true.
Arena.ai is supposed to factor for that.

And you still have to scroll down a long way.
It really depends on the task, asking general questions hasn’t changed much since ye olde days of GPT-4.
But for coding or reasoning tasks the difference is very noticeable.
Ah I think it depends on your needs. As I say I only ask about one question a week and they tend to be broad ones (often historical) that I think I’m unlikely to get a hallucination from and I feel like more usage is likely to get them more funding. I’ve removed ChatGPT as a result.
Interesting graph though, no question. Maybe they’re the Lemmy of AI :)
imHo, idea of an ai is not so bad on itself.
What sux, is that idiots replaces own brains for ai.
What sux more, that we don’t have any law to register ai.
And what super sux, that big tech ai uses lack of law to control brainless idiot masses.
Also privacy, stealing content and art for learning, bugs, still bit early versions of ai, that rarely leads to harm.
If big tech do not own all by that, maybe in 5-10years, ai will be great thing. For now, just a silly toy to predict weather no less accurate than other forecasts, maybe to tldr wiki articles.
For Mistral, once did read, this one is already controlled by some extra laws, making it fairly nice? No idea if still up to date or just rumour.
Don’t forget that we don’t actually price energy use in a way that covers all the downstream environmental harms, and it’s not like AI has lead to fewer people, so right now we’re just ballooning our per capita energy (and consequent environmental harm), massively.
I’ve been using the pro version for translations and editing. It’s good, and it’s European.
Does it translate better than DeepL? (German)
Oh yes! I already disliked deepl, because I had security issues on my windows and MacBook; deepl installs some weird shit in the background that my setup didn’t like (it turned out it is hard to remove too). But, mistrals vibe is definitely better at grammar and translations (I use it for 5, 6 languages). I also use it to write contracts for business partners - I’m quite a fan of AI actually.
I haven’t used Mistral for a while now, because I have been using Lumo that has been good enough for my once-in-a-while questions, but I’ll give Mistral another try next time i have a question.
I used to subscribe to it personally but the quality of work was just not there compared to ChatGPT, Claude, or even Gemini (all three of which my client’s company pays for me to use; I wouldn’t spend a penny on US big tech). I do have Ministral running on my computer to do more simple tasks, though!
Mistral markets itself not as frontier but as custom-integratable with company data and knowledge bases.
I tried using it about a year ago and it couldn’t even make a table from an evaluated formula without being consistently wrong. It may be different now, but it was failing quite basic algebra.
Trusting an LLM for algebra seems like asking for trouble to me. Give them a tool to perform the calculations instead.
it was failing quite basic algebra.
I think a lot of the LLM’s were back then. It has definitely introduced a “thinking” element now so might be worth another whirl.
I’ve been using Mistral for a while now, and medium-latest has been quite good in terms of responses. But it seems Medium is being deprecated in April in favor of the newer Mistral Large, which, yes, is good, but incredibly slow because of the thinking parameter.
I’m hoping that’s mostly due to how new the model is and that it’ll become more performant by April.
Black Forest Labs is my favorite European AI company, but they don’t really do text.




