Open Weights Aren’t Just Chinese Anymore - DTNS 5368

Tom Merritt (0:08)

Oh, this is the Daily Tech News for Tuesday, October 6th, 2026 We tell you what you need to know, give you important context, and we're trying our best. We're like the shepherd, helping each other understand.

Jason Howell (0:21)

We'll try our very, very best. Today, France's, hey, Mistral, Mistral, over here.

France's Mistral, I don't even know what that's from, but it's in my head, so there you go.

Tom Merritt (0:32)

It's film noir in general, as far as I know, yeah.

Jason Howell (0:35)

I suppose so. Maybe it's applicable, I don't know. But France's Mistral finds a third way between open and closed.

Tom Merritt (0:42)

What could that be? We'll talk about that. Partially open. I'm Tom Merritt.

Jason Howell (0:47)

I'm Jason Howell.

Tom Merritt (0:47)

Let's start with what you need to know with that partially open big story.

So, yes, Mistral released a public preview of Mistral Large 4 It's the fourth in the series of Mistral Large. You can just call it ML4, or if you're on very familiar terms with this multimodal model, you can call it Le Chonk, as in Chonky, as in it's got a trillion parameters. Mistral says it was trained on 3,800 Nvidia Grace Blackwell GPUs. So not Chonky on the compute side. They didn't have to go anywhere else. They're like, we just used our own compute.

It says it is two to three times fewer than the amount of GPUs used by all of the competing Chinese companies that release open weight models. ML4, or Le Chonk, can be accessed now through a public preview API in Mistral Studio, and then after safety testing is complete, Mistral plans to make it an open weight model. Somewhere around October 27th, look for that. Mistral expects it to outperform other open weight models and possibly outperform some closed frontier models in specific areas. They say this ML4 is good for security, for finance, and for chip design, which I guess kind of makes sense, since two of Mistral's main backers are ASML, which makes the lithographic machines for everyone, and Samsung, which makes chips.

It is at a trillion parameters, you might have guessed, too big to run on your desktop. But once the weights are released, companies will be able to run it on-prem, or in a private cloud. It might even be too big, according to TechCrunch, to run in a university, because it's pretty darn big. But for companies that have the space for it, you'll be able to run it yourself and not worry about giving that model over to Mistral or anybody else. Which I think what Mistral is targeting is like, hey, you want sovereign AI companies, but you don't want to rely on China. How about France? We're friendly most of the time, at least in AI.

Jason Howell (2:52)

Well, yeah, we're friendly most of the time. And I think to a certain subset of people who might consider this, we're also not the US, right?

Tom Merritt (2:59)

We're not the US or China, right? We're the third way.

Jason Howell (3:03)

Exactly. And so maybe that's part of the midway between open and closed, is not the US and China.

Anyways, I love LeChonk.

I know it's stupid to start with that, but I picture like a character with LeChonk. I think that was a good move. I just think Mistral, a lot of what Mistral is doing is really smart. I think they're playing an important role at a time where we're so used to hearing about these frontier releases from US companies. And then on the other side, it's the Chinese open weight and everything happening with open weight models over there and everything, but attached to China. And some people look at that and they say immediately like, well, then I'm not going to go there. Having a company like Mistral that is coming from Europe, and Europe has for a while now, had an AI company that is doing good things and going toe to toe with in that conversation with everyone else. So I think it's a good place for them to be. And I mean, a trillion parameter model, that's nothing to sneeze at. This sounds impressive.

Tom Merritt (4:12)

It's chunky.

Jason Howell (4:13)

I get it.

Tom Merritt (4:13)

Chunky.

Jason Howell (4:13)

Very chunky. It's a chunker.

Tom Merritt (4:15)

Mistral is the Spotify of European AI. You know, it is the one you can point to and say, but they're doing well.

I know some people in the industry argue about whether they're given more credit than they're due simply because they're Europe's representative in the fight. But I think this is a chance for them to say, no, that's, there should not be truth to that. You know, they're gonna say, this is a very competitive model. Is it better than OpenAI or is it better than Claude? Maybe not. But it is designed to be very good for other things. And, you know, they're the Avis. We try harder of the situation. You know, and that's what we want. We want competition on a global scale for this sort of thing. So I'm glad to see that some of the speculation that Mistral was just kind of fade into being a service provider, providing access to open weight models to other people, turned out, at least for the moment, not to be true. I'm glad about that.