Latest / DX Today | No-Hype Podcast & News About AI & DX / Mistral Gives Away a Trillion Parameter Model for Cyber Defense

Transcript

Mistral Gives Away a Trillion Parameter Model for Cyber Defense · DX Today | No-Hype Podcast & News About AI & DX

• 0:00
  SPEAKER_01: Welcome to the DX Today Podcast, your daily deep dive into the AI ecosystem.
• 0:05
  SPEAKER_01: I'm Brian, and joining me as always is Jessica.
• 0:08
  SPEAKER_00: Thanks, Brian.
• 0:08
  SPEAKER_00: And today we have a story that I think gets at one of the biggest fault lines in the whole industry, which is who actually gets to own and control the most capable AI models.
• 0:17
  SPEAKER_01: Right.
• 0:17
  SPEAKER_01: So on Thursday, October 29th, the French company Mistral released the full open weights for Mistral Large 4, which is a model with roughly one trillion parameters that anyone can now download.
• 0:29
  SPEAKER_00: And the internal nickname is LeConc, which I love because it is honest.
• 0:34
  SPEAKER_00: This thing is enormous.
• 0:36
  SPEAKER_00: It was first shown in a public preview on October 6th, and now the actual weights are sitting on Hugging Face.
• 0:41
  SPEAKER_01: Before we get into the politics, let me do my usual thing and ask what this actually is, because a trillion parameters sounds impressive, but I want to know what it means for somebody running real systems.
• 0:52
  SPEAKER_00: Fair question.
• 0:53
  SPEAKER_00: It is a mixture of experts model, so out of that trillion total parameters, only about 52 billion are active for any given token, according to Ms.
• 1:02
  SPEAKER_00: Strahl, which keeps the compute per word much lower than the size suggests.
• 1:06
  SPEAKER_01: Okay.
• 1:06
  SPEAKER_01: So think of it like a huge hospital with hundreds of specialists on staff, but any single patient only sees a handful of them.
• 1:14
  SPEAKER_01: You pay for the building, but each visit is relatively cheap.
• 1:17
  SPEAKER_00: That is a great way to put it.
• 1:18
  SPEAKER_00: And the building part matters because you still have to hold every one of those specialists in memory to run the model, even if most of them are idle on a given request.
• 1:27
  SPEAKER_01: So what is the real footprint?
• 1:29
  SPEAKER_01: If my team wanted to bring this in-house tomorrow morning, how much hardware are we talking about before we can even say hello to it?
• 1:35
  SPEAKER_00: The standard compressed checkpoint at 8-bit precision comes in at roughly 1.1 terabytes.
• 1:42
  SPEAKER_00: There is also a smaller 4-bit format aimed at the newest Nvidia Blackwell chips, and fine-tuning reportedly fits on a single server node with 8 high-end GPS.
• 1:51
  SPEAKER_01: So this is not something you run on a laptop in the break room.
• 1:55
  SPEAKER_01: This is a serious data center commitment, even if it is cheaper than training anything yourself from scratch.
• 2:01
  SPEAKER_00: Mostly true, but the community is already squeezing it.
• 2:05
  SPEAKER_00: A group called Unsloth released an extreme one-bit version that shrinks it to about 223 gigabytes, and they say it keeps about 91% of the top answer accuracy.
• 2:15
  SPEAKER_01: I will believe the 91% when I see it hold up on my own workloads.
• 2:20
  SPEAKER_01: But I take the point.
• 2:21
  SPEAKER_01: The open release means other people get to experiment with it in ways Mistral never would have done itself.
• 2:27
  SPEAKER_00: Exactly.
• 2:28
  SPEAKER_00: And that is the whole philosophy.
• 2:30
  SPEAKER_00: It also launched with support on day one in the popular serving tools like VLLM and SG Lang.
• 2:37
  SPEAKER_00: And it runs on AMD chips and Google Tensor processors, not only on NVIDIA hardware.
• 2:42
  SPEAKER_01: Let us talk about whether it is actually any good, because every launch announcement claims to be state-of-the-art, and I have learned to read those benchmark tables with a fairly suspicious eye.
• 2:52
  SPEAKER_00: As you should.
• 2:53
  SPEAKER_00: And I want to be clear that these numbers are reported by Mistral and its partners, not independently verified yet.
• 2:59
  SPEAKER_00: But on a software engineering test called Deep SWE, it scores about 61.7%.
• 3:05
  SPEAKER_01: And how does that stack up against the other open models, especially the Chinese ones, which have honestly been leading the open weight world for most of this year?
• 3:13
  SPEAKER_00: That is the honest part of the picture.
• 3:15
  SPEAKER_00: It roughly ties GLM 5.3, the model from the Chinese lab Jeepu, but it still trails Kimik 3 and the newest Deep Seek Flash release on that particular coding test.
• 3:27
  SPEAKER_01: So it is competitive but not dominant.
• 3:29
  SPEAKER_01: That actually makes me trust the story more, because a company that admits it is not first on every chart is probably not inflating everything else either.
• 3:37
  SPEAKER_00: And there are areas where it does lead.
• 3:39
  SPEAKER_00: On a hard coding benchmark called Frontier Code 2.0, it posts the top open weight score.
• 3:45
  SPEAKER_00: And on a terminal agent test, it climbed from 28% to 41% after extra reinforcement training.
• 3:51
  SPEAKER_01: What I like there is the efficiency claim.
• 3:54
  SPEAKER_01: Mistral says it hits those agent scores while using about 47% fewer tokens per task than the comparable Chinese model, which translates directly into lower bills.
• 4:04
  SPEAKER_00: And speaking of bills, the price is aggressive.
• 4:06
  SPEAKER_00: On the Mistrel platform, it costs 70 cents per million input tokens and$2.10 per million output tokens, with cashed input dropping to just seven cents.
• 4:16
  SPEAKER_01: That is a very low price for a model in this weight class.
• 4:19
  SPEAKER_01: So who is the real target here?
• 4:21
  SPEAKER_01: Because it does not sound like they are chasing consumers who want a chatbot on their phone.
• 4:25
  SPEAKER_00: The target is very clearly cybersecurity, and specifically governments and big enterprises in Europe.
• 4:31
  SPEAKER_00: Mistral is pitching this as sovereign cyber defense, meaning tools to detect, analyze, and respond to threats across code and infrastructure that you control yourself.
• 4:40
  SPEAKER_01: Here is where it gets interesting for me.
• 4:43
  SPEAKER_01: The same week, Google said its new security-focused model, Gemini 4 Carbon, would go first only to a vetted group of trusted cyber defenders, with the United States government already testing it.
• 4:55
  SPEAKER_00: Right.
• 4:56
  SPEAKER_00: Two completely opposite philosophies landing in the same news cycle.
• 5:00
  SPEAKER_00: One company says this capability is powerful enough that we will handpick who gets it, and the other says, here are the weights.
• 5:07
  SPEAKER_00: Take them home.
• 5:08
  SPEAKER_01: And I can see the argument for both.
• 5:10
  SPEAKER_01: If a model is very good at finding and exploiting software vulnerabilities, giving it to everyone on the internet does make me a little nervous, as someone who has been paged at three in the morning.
• 5:19
  SPEAKER_00: Mistral has an answer to that, though.
• 5:21
  SPEAKER_00: Their vice president of science, Pierre Stock, argued that open weights mean access can never be revoked by a provider.
• 5:28
  SPEAKER_00: And that defensive capability needs to reach everyone quickly because attackers will get similar tools anyway.
• 5:34
  SPEAKER_01: That last part is the crux, is it not?
• 5:37
  SPEAKER_01: The attackers are not waiting for permission.
• 5:40
  SPEAKER_01: The question is whether openly releasing the model helps defenders more than it helps the people they are defending against.
• 5:46
  SPEAKER_00: And Mistral made a pointed claim here.
• 5:49
  SPEAKER_00: On its own reproduce and patch test, where a model has to recreate a vulnerability and then fix it, the large American closed models scored near zero, not because they lacked skill, but because they refused the task.
• 6:01
  SPEAKER_01: That is a fascinating detail, and I think it is the most practical point in this whole story.
• 6:07
  SPEAKER_01: If you run a security team and your tool refuses to look at the exploit, it does not matter how smart it is.
• 6:13
  SPEAKER_00: Exactly.
• 6:14
  SPEAKER_00: Mistral says those provider-level refusals can block legitimate vulnerability research and incident response.
• 6:20
  SPEAKER_00: Their model scored about 82% on that test and 93% on a capture-the-flag style benchmark called Cybench.
• 6:27
  SPEAKER_01: But then I have to ask the obvious follow-up.
• 6:31
  SPEAKER_01: If it is willing to do the scary work for defenders, what stops it from doing the same scary work for a criminal who downloads it from Hugging Face this afternoon?
• 6:39
  SPEAKER_00: Mistral reports a 96.6% average refusal rate on clearly malicious cyber requests across three public safety benchmarks, and it says it red-teamed the model with security firms, vetted partners, and state authorities before release.
• 6:53
  SPEAKER_01: And of course, once the weights are out, anyone can fine-tune those refusals away.
• 6:58
  SPEAKER_01: So I would treat that safety number as a description of the default behavior, not a guarantee about what will be running in the wild next month.
• 7:06
  SPEAKER_00: That is fair, and I will not pretend otherwise.
• 7:09
  SPEAKER_00: But I would push back a little because the same is true of every strong open model already out there, including the Chinese ones that nobody in Europe controls at all.
• 7:18
  SPEAKER_01: That is a good point, and it reframes the decision.
• 7:22
  SPEAKER_01: The choice for a European government may not be open versus closed.
• 7:26
  SPEAKER_01: It may be a European open model versus a Chinese open model versus an American closed one.
• 7:31
  SPEAKER_00: And that is where the money and the politics come in.
• 7:34
  SPEAKER_00: This model was trained from scratch on about 3,800 NVIDIA Grace Blackwell chips sitting in the data centers that Mistral runs itself in Europe, which is a big deal for sovereignty.
• 7:44
  SPEAKER_01: So the training, the hosting, and the company are all on European soil.
• 7:49
  SPEAKER_01: For a defense ministry or a bank regulator, that is not a marketing slogan.
• 7:54
  SPEAKER_01: That is a procurement requirement they actually have to meet.
• 7:57
  SPEAKER_00: Their tagline is literally forged in Europe, built for AI sovereignty.
• 8:02
  SPEAKER_00: And Chief Executive Arthur Mensch describes the goal as making AI a tool for empowerment, not dominance, which is a not very subtle jab at the American giants.
• 8:10
  SPEAKER_01: Let me play skeptic on the business side.
• 8:12
  SPEAKER_01: Giving away your best model for free seems like an odd way to build a company that just raised a lot of money from investors who expect a return.
• 8:20
  SPEAKER_00: It is not quite free.
• 8:21
  SPEAKER_00: And the license is where the strategy shows.
• 8:24
  SPEAKER_00: The earlier version, Mistral Large 3, used the very permissive Apache license.
• 8:29
  SPEAKER_00: This one uses what Mistrel calls a modified MIT license with an important catch.
• 8:34
  SPEAKER_01: Let me guess where the catch is.
• 8:36
  SPEAKER_01: It is the cloud providers who would like to host it and resell it without sending a single euro back to Paris.
• 8:41
  SPEAKER_00: Exactly right.
• 8:43
  SPEAKER_00: Internal use, modification, even reselling products built on it are allowed.
• 8:47
  SPEAKER_00: But anyone running it as a model, as a service business, needs a separate commercial license from Mistral before they can make money hosting it.
• 8:54
  SPEAKER_01: That is clever, and honestly, it is pragmatic.
• 8:58
  SPEAKER_01: You let enterprises and governments run it for themselves, which builds trust, while blocking the big hosting companies from simply cloning your product and undercutting you on price.
• 9:07
  SPEAKER_00: And it ties into their funding.
• 9:09
  SPEAKER_00: Mistral links this work to a 3 billion Euro Series D round, which it calls the largest equity round ever raised by a European technology company, so investors are clearly betting on sovereignty as a business.
• 9:20
  SPEAKER_01: I do wonder about the ongoing cost of keeping up, though.
• 9:23
  SPEAKER_01: Releasing one great model is a moment, but staying near the frontier every few months requires a level of compute spending that only a handful of companies can sustain.
• 9:33
  SPEAKER_00: That is the real test, and Mistral seems aware of it.
• 9:36
  SPEAKER_00: They say their reinforcement learning pipeline alone runs on about 3,000 GPUs and generates roughly 33 billion tokens of training data every single day.
• 9:45
  SPEAKER_01: That is a serious operation for a company that most people outside the industry have never heard of.
• 9:51
  SPEAKER_01: It tells me they are not treating this as a one-time stunt, but as a continuing program.
• 9:56
  SPEAKER_00: And the multimodal side helps their case with government buyers.
• 9:59
  SPEAKER_00: The model reads documents, charts, technical drawings, and even satellite imagery.
• 10:04
  SPEAKER_00: And it supports more than 160 languages, including every official language of the European Union.
• 10:10
  SPEAKER_01: So, what should a technology leader listening to this actually do on Monday morning?
• 10:14
  SPEAKER_01: Because I think a lot of people hear trillion parameters and assume this has nothing to do with them.
• 10:19
  SPEAKER_00: My first suggestion is to look at your security workflows and ask where a closed model has refused to help you.
• 10:25
  SPEAKER_00: If that friction is real, a self-hosted model like this is now a credible option worth piloting.
• 10:30
  SPEAKER_01: And my second suggestion would be to price out the hardware honestly.
• 10:35
  SPEAKER_01: Most teams will not run the full model themselves, so the realistic path is a trusted regional host, and the license makes that a conversation with Mistral directly.
• 10:44
  SPEAKER_00: And third, if you work anywhere near regulated industries in Europe, keep a very close eye on this, because a credible homegrown option changes the negotiating position against every American vendor at the table.
• 10:55
  SPEAKER_01: I will add one more from the skeptic corner.
• 10:58
  SPEAKER_01: Wait for independent benchmarks before you rewrite your roadmap, because every number we discussed today came from Mistrol or its close partners, and the real test is your own data.
• 11:09
  SPEAKER_00: Agreed.
• 11:09
  SPEAKER_00: And that is the bigger story here.
• 11:12
  SPEAKER_00: The frontier is no longer only a two-country race between the United States and China, and Europe just put a trillion parameter stake in the ground that anyone can download.
• 11:21
  SPEAKER_01: Whether that makes the world safer or riskier is going to depend on who picks it up first, the defenders or the attackers.
• 11:27
  SPEAKER_01: And I suspect we will be talking about that question for a long time.
• 11:31
  SPEAKER_00: Which is exactly why it matters for everyone building with AI right now.
• 11:35
  SPEAKER_00: Because the choice between open and closed models is no longer a philosophical debate.
• 11:39
  SPEAKER_00: It is a practical decision showing up in budgets and procurement.
• 11:43
  SPEAKER_01: That's all for today's episode of the DX Today podcast.
• 11:46
  SPEAKER_01: Thanks for listening, and we'll see you next time.