Mistral Drops New AI Models for Laptops and Phones "Les Ministraux
10:01 Share

We have great news from Mistral AI, as they just released their first model specifically designed to run on computers and phones, or what we call edge cases. This is a really fascinating development from this Paris-based startup.

If you are interested

in making money with AI tools. I would love to have you as a member of the AI Hustle School community. It's a community where every week I post exclusive content showing you how I scale my current businesses with AI tools, how you can do it with all the processes, workflows, and tools I use, and how I make money with AI tools different from SideHustle. So if that's something that interests you, we have an amazing community, exclusive content

and teach you how to grow your business, join the AI Hustle School community and we'd love to have you involved in the process. Anyway, let's start by seeing what Mistral is currently doing.

This is all very interesting. It's a new family of models that Mistral calls "Les Ministraux". And essentially, they can be used or returned for a bunch of different use cases. They obviously have basic text generation, but they also have other interesting things. In particular, they say you should use them with some of their other... their larger, more capable

So you use this hybrid of a smaller model that's more refined with other better models. And some people, I know they're going to be skeptical of this approach because they're like, "Why would I do that?" "I'll just use the best model. I'll just use GPT-4O preview to just get what I want. I just want the best model." So this is really very useful for developers, myself included. I'm currently developing a number of different AI applications and specifically with AIbox,

I'm building right now, we use smaller models when people query

we use smaller models to determine which model to query. It's much more efficient than just using something like GPT-4.0 and it's also capable. There are a lot of things that developers use. Some of these smaller models are faster, more nimble. And of course, as they mention here, you can do it directly on your phone and on your phone, which is a big benefit. This is specifically

privacy and security. You're not sending this up to a cloud, no one can intercept it, no one can stop it. And so I think that for a lot of people in a lot of places, it's something that's very interesting to them. So like I mentioned, there are two models available, Ministral 3B and Ministral 8B. Both of them have a context window of 128,000 tokens. So this means that they can both take in about the length of a 50-page book. This is...

This is fantastic. That's as large as you need for most use cases. You know, 99% of users, this is perfect. So overall, I'm really excited about what they've done here. They had a blog post that broke down everything that's going on here. And they had an interesting quote that said, "Our most innovative clients and partners have increasingly demanded local private inference for critical applications such as on-device translation, internet lists, smart assistance, local analytics, and autonomous robotics."

Les Ministraux were built to offer an efficient and low-latency competitive solution for these scenarios. From independent hobbyists to global manufacturing teams, Les Ministraux deliver a wide variety of use cases. I'm really excited about this. They've already triggered a bunch of interesting use cases here, on device translation. You can imagine you're in a foreign country, maybe you don't have good cell reception, maybe you don't have a phone plan there, or you don't have internet.

It can translate directly on your phone and help you communicate with people. That's another incredible use case. The other one is smart internet assistance. So again, you have this situation where you're putting yourself in the mountains on a hike and you don't have good reception because you don't have Starlink connected to your phone, I'm sure that problem will go away in the next 5 years, but today, and you're getting around and you want to know how...

for

for all sorts of things, but also, you can think of many military applications that this could have. It's a terrifying prospect, but you can see where it's going. And then, they also say that it's essentially built to offer compute efficiency. So it doesn't need to do crazy computer games. That is, it will be directly on your phone and then, the latency is low. So these things are going to be incredibly fast. You're not sitting around waiting. They're very fast. And that's what we see with GPT-40 Mini as well, is that it's

very, very fast. So Global Manufactures and a bunch of other people are using this. On their blog post, they go through the benchmarks. The thing is pretty robust in terms of the pressure of different benchmarks and what it can accomplish. And...

I have to give them props because at the end of their blog post, they put a meme on it that says 2023 and then they say "Mistral 7B" and it's like all the Ninja Turtles, their kids with their grandmaster in the middle. And so it's like, "OK, last year we had one thing and then it says 2024, it's all the aged Ninja Turtles and their grandmasters are now smaller than them, but each of the Ninja Turtles is labeled

Ministral 3B, Ministral 8B, PixTrail, Ministral Small. I think they're stating that there's been something interesting in the last year and there's been a big decline this year. And again, it has to be remembered that Mistral does a lot of things in open source. It's an absolutely fascinating company that really does a lot. The thing to break down here is that Ministral 8B is available for download. You can go find it today. I'll give them a pat on the back for that.

But I would say that it's strictly for research purposes. If you want your devs or your company to use this, either their 8B or their 3B Ministral, and you want to do development or deployment steps, you need to contact them for a commercial license.

You can do that, but it's going to be a contact thing. And obviously, that's probably part of their way of planning to make money. So otherwise, developers can use Ministral 3B and 8B through their cloud platform, which is called La Platform. And they also have other clouds that they've partnered with in the coming weeks, that they're going to have on them as well. So that's going to be interesting. I'm speculating maybe Microsoft Azure or maybe Google.

or maybe even Amazon, AWS, so it could be all or some of those that they'll need the application on. Right now, the Ministral 8B costs about 10 cents per million outputs, so that's about 750,000 words, very, incredibly well-priced, and their Ministral 3B costs 4 cents per million outputs.

It's incredible. For 10 cents, you can generate a quarter of a million words or on their smaller model, 4 cents can generate a quarter of a million words. That's...

They're not the only ones doing this. There's certainly been a trend towards smaller models that are cheaper and faster to process or finish. I've crunched a bunch of use cases, but really, some of these things can be quite expensive. Google has added a bunch of new models to their Gemini family. Microsoft has Pi.

in their collection of models. And recently, the Llama suite introduced a bunch of new models that are really optimized for edge hardware as well. So a lot of people have come up with this thing. Mistral claims that Mistral 3B and Mistral 8B will, you know, defend Llama, Gemini, as well as their own Mistral 7B on a bunch of different AI benchmarks that they do. So for me, this is very interesting.

All in all, very impressed with this company. We recently saw that they raised $640 million. They even launched a free tier for developers to test their AI models. And they also launched a bunch of new services and an SDK to essentially let their clients fine-tune their models. So a lot of impressive things coming out of Mistral. A very innovative company. They announced their PIX-Trail 12B. Just a lot of impressive things. And of course,

...

and prove everyone wrong. I love a company with a chip on its shoulder. Mistral is certainly that company. This is one I'm going to keep my eye on going forward. Again, if you're interested in scaling and growing your business with AI tools, I'd love for you to join the AI Hustle School community or if you're trying to start an AI side hustle and make money on the side, I'd love to cover how I was able to make over $12,000 in several months by uploading free videos to Amazon and a

Subtitling ST' 501