Paris-based Mistral wanted to develop a top-tier AI model to rival OpenAI and Anthropic. That didn’t work out. But it turns out lots of folks don’t care if the AI is bleeding edge – as long as it wasn’t made in America or China.
… [Cofounder and CEO of Mistral Arthur] Mensch’s vision for Mistral, and AI itself, can be summed up in one word: independence. Unlike its black-box Silicon Valley rivals, most of Mistral’s AI models are what techies call “open weight.” In this sort of open-source model, customers are free to get under the hood, customize the AI using their own data or download it for free to run offline (or from a laptop).
The message resonates. Old-school execs are spooked by the world-consuming rhetoric of OpenAI and Anthropic and the emerging threat of Chinese AI companies. Mensch’s talk of control and sovereignty is soothing, as is his pitch that Mistral will deploy engineers to set up and run the tech for them. Your data doesn’t even need to leave the office, let alone the country.
“We are really the only company that allows [building] core business automation and products on top of an open stack, and that is something that is valuable everywhere in the world,” says Mensch, 33, from Mistral’s offices in the trendy 10th arrondissement of Paris, as kids play soccer in the courtyard out back.
…
A bit short sighted to say their goals didn’t work out.
Look at what they’ve built: STT and TTS pipelines, multilingual models, document retrieval, parsing, and summarization.
Those are enterprise and government use cases, which makes sense for a European sovereign AI initiative aimed at governments.
They do also have le chat, which is like ChatGPT. devstral is currently quite behind, but also not reinforcement learned on benchmarks to shit like Qwen/GLM models are.
Mistral still sit in a good spot of small/hostable, private, and good world knowledge plus communication.
On the frontier level they’re probably 8-12 months behind, though devstral was the first local agent I could run.
Yes, the models are open, the API’s are European, and the outomes are good enough. My money is better spent on Mistral than on the American ones; they can have their bleeding-edge closed stuff.
When they can say that their electricity comes for either nuclear or renewables, I’ll consider subscribing.
What’s great about Mistral is that they have to follow the EU AI act.
While it may not compare to the closed, corporate “frontier” models that are running each query on racks upon racks of servers worth millions of dollars in a corporate data center, it’s not clear to me that’s a reasonable comparison. Mistral is actually one of my personal favorite models and it has nothing to do with being French (although that is a huge bonus, it doesn’t really contribute to my method of scoring and ranking of different models). It’s simply more reliable and consistent than a lot of the other open-weight models, it follows directions well and has good task comprehension. Even with abliteration (guardrail removal) it doesn’t generally have weird artifacts and it uses diverse, clear and reasonable language in its responses. It also handles Markdown really nicely in a very consistent way without needing guidance. To me, these are more important properties than simply “containing the most recent training data” or how many agents it has. At the end of the day I’m going to judge a model based on its output more than how it gets there and it turns out you don’t need all the latest whiz-bang features to get acceptable output depending on what exactly you’re doing with it.
It may not be as technically impressive as even the latest greatest open-weight models like Qwen, but it’s still a great model whether it’s French or not, and I would still use it just as often as I already do even if it were Chinese or North Korean or Russian I don’t care, all that matters is that it’s open-weight, runs with reasonable hardware requirements, is flexible in how it can be used, and produces good output. Mistral ticks all the boxes.
deleted by creator
Le Chat is the LLM I use (if I actually want to use AI). It does the job.
I haven’t checked the competition to be honest. Mistral is good enough for my needs.
Are others that much better?The name alone takes the cake “le chat”
What’s really important IMO is if uses stolen data/will it steal mine, and if it will burn a box of seal puppies to basically do a Google search.
Not being american does not mean that this “AI” is better or worse than all of the other questionable “AI” out there.
They are better than copilot but worse than gemini and chatgpt.
The app closes everytime I open it and says some google srvices error. It doesnt work if you get it without the playstore.
I really want it to succeed but it needs to keep going.
So for me it works fine on e/OS doenloaded through Aurora Store.
The results are not as good as the other huge LLMs, but for the stuff I use it is efficient enough.
Aurora would have google services, I must have got an apk or something.
I just wish all apps offered direct apk downloads without the google mandated shit
