# Why Closed AI Providers Block Cyber Defense | Sagar Vaze

Markdown copy of https://realalpha.app/v/gJZLJSIEbtQ/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy.

Sagar Vaze argues that open-weight models let companies customize AI and give cyber defenders an alternative when closed providers block requests.

- Show: MTS
- Hosts: [Theo Jaffee](https://realalpha.app/people/theo-jaffee/) & [Sophia Dew](https://realalpha.app/people/sophia-dew/)
- Guest: [Sagar Vaze](https://realalpha.app/people/sagar-vaze/) is a research scientist at Mistral, leading the multimodal team.
- Published: Oct 29, 2026
- Length: 21 min

## The best minutes

- [14:05](https://realalpha.app/v/gJZLJSIEbtQ/?t=845) [Closed-model refusals can leave cyber defenders without recourse](https://realalpha.app/s/closed-model-refusals-can-leave-cyber-defenders-without/) (3 min)
- [9:52](https://realalpha.app/v/gJZLJSIEbtQ/?t=592) [Mistral trained its trillion-parameter model on its own compute](https://realalpha.app/s/mistral-trained-its-trillion-parameter-model-on-its-own/) (3 min)
- [3:26](https://realalpha.app/v/gJZLJSIEbtQ/?t=206) [Fine-tuning can beat prompting on narrow enterprise tasks](https://realalpha.app/s/fine-tuning-can-beat-prompting-on-narrow-enterprise-tasks/) (2 min)

## Also worth watching

- [5:45](https://realalpha.app/v/gJZLJSIEbtQ/?t=345) [Compute is Mistral's main gap to the leading AI labs](https://realalpha.app/s/compute-is-mistrals-main-gap-to-the-leading-ai-labs/) (2 min)
- [1:01](https://realalpha.app/v/gJZLJSIEbtQ/?t=61) [Open weights let firms turn proprietary expertise into a moat](https://realalpha.app/s/open-weights-let-firms-turn-proprietary-expertise-into-a/) (2 min)
- [16:44](https://realalpha.app/v/gJZLJSIEbtQ/?t=1004) [Mistral trained multimodal capabilities from the start](https://realalpha.app/s/mistral-trained-multimodal-capabilities-from-the-start/) (2 min)

Skip the cat-meme discussion and the closing sponsor reads, both about two minutes.

## Summary

Sagar Vaze is a research scientist at Mistral and leads its multimodal team. He worked on Mistral Large 4, an open-weight model trained from scratch on Mistral’s own compute near Paris.

Vaze argues that closed providers’ cyber safeguards can block defensive requests, making capable open-weight models important for enterprises protecting their infrastructure. He says companies can fine-tune open models on proprietary data to build a competitive advantage, and that fine-tuning remains useful even as models improve. Vaze says Mistral needs more compute to close the gap with frontier labs and describes Large 4 as a trillion-parameter model trained on 1E25 FLOPS. He also discusses on-policy distillation for different reasoning modes and training the model multimodally from the start.

### Minute by minute

- [0:00](https://realalpha.app/v/gJZLJSIEbtQ/?t=0) **Vaze says open weights give defenders recourse.** Vaze says providers can refuse defensive cyber requests because they resemble offensive ones to the classifiers used to screen them. He argues that open-weight models of sufficient strength give enterprises a way to scale up their defenses.
- [1:01](https://realalpha.app/v/gJZLJSIEbtQ/?t=61) **Open weights widen access.** Vaze says open weights have been a core Mistral principle since the company began. He argues that AI should not be controlled by a few providers that can decide who gets access or turn models off.
- [2:17](https://realalpha.app/v/gJZLJSIEbtQ/?t=137) **Customization can preserve enterprise advantage.** Vaze says Mistral can build a business around customization and fine-tuning. In Mistral’s vision, companies can adapt an open-weight model to proprietary data and expertise, turning legacy knowledge into a competitive advantage.
- [3:26](https://realalpha.app/v/gJZLJSIEbtQ/?t=206) **Fine-tuning still helps narrow tasks.** Vaze says a small fine-tune on collected data beat prompt-tuning strategies for a simple internal classification task, even with a strong 30b model. He also says Mistral Small from Jan 2025, fine-tuned on company-specific programming languages, can perform as well as some of the best closed models.
- [5:45](https://realalpha.app/v/gJZLJSIEbtQ/?t=345) **Vaze says compute is key to catching up.** Vaze says Mistral could have been at the frontier sooner if it had been founded in 2019 or 2020, rather than 2023. He says compute is the key to closing the gap and that Mistral had raised 6 billion, though he could not recall whether dollars or euros.
- [10:02](https://realalpha.app/v/gJZLJSIEbtQ/?t=602) **Large 4 scales to a trillion parameters.** Vaze says Mistral Large 4 has 1 trillion parameters and was trained on 1E25 FLOPS using compute at Mistral’s data centers near Paris. He calls it Mistral’s best agentic coding model and says it tied with GLM 5.3 on the AA coding index while using 50% less output tokens.
- [11:58](https://realalpha.app/v/gJZLJSIEbtQ/?t=718) **Open releases support shared progress.** Vaze says open releases let researchers learn from one another. He cites Mistral’s Mixtral 8X7B, which he calls the first open-weight mixture-of-experts model, and says DeepSeek V3 influenced Mistral’s later ML3 architecture.
- [12:56](https://realalpha.app/v/gJZLJSIEbtQ/?t=776) **Distillation supports different reasoning modes.** Vaze says reinforcement learning produces low, medium, and high reasoning checkpoints with different token use and performance. On-policy distillation combines them into one student model that can use each mode.
- [14:16](https://realalpha.app/v/gJZLJSIEbtQ/?t=856) **Vaze links attacks to closed-weight models.** Vaze says significant malicious attacks have been carried out using closed-weight models that were jailbroken or escaped their sandbox during training. He says provider classifiers can block defensive requests that resemble offensive ones, leaving some enterprises with no recourse but open-weight models.
- [16:37](https://realalpha.app/v/gJZLJSIEbtQ/?t=997) **Large 4 trains multimodally from scratch.** Vaze says images were part of Large 4’s training from the start, rather than added through later fine-tuning. He says the model handles images and video, improved at 3D asset generation, and required research to build multimodal capability without regressing on coding and other text tasks.

## Transcript

MTS, hosted by Theo Jaffee and Sophia Dew, with Sagar Vaze.

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

[0:00](https://realalpha.app/v/gJZLJSIEbtQ/?t=0) And on the other hand, we've got defenders, you know, enterprises trying to shore up their vulnerabilities, their own defenses. We're finding they're having their requests being refused by the same providers for completely understandable reasons, because in practice, these providers have classifiers under the hood, which say, is this a malicious cyber request or is this a regular request? And for the cutting edge defense capabilities, these defensive requests in practice look quite similar to the offensive ones. The net-net effect of that is that people trying to shore up their defenses from malicious attacks, often from closed-weights models, are finding that they have no recourse but to use an open-weights model.

[0:38](https://realalpha.app/v/gJZLJSIEbtQ/?t=38) And so we think that it's critical that there are open-weights models of the same strength, of sufficient strength, that people can really scalably defend their infrastructure. We are back. We are live with Sagar Vaze, who's a research scientist at Mistral, leading the multimodal team. He worked on Mistral Large V4, which was just open weights, released as an open weight model today, which is very exciting. Samara, welcome. Thanks for having me. Happy to be here. Absolutely. So tell us about, why does Mistral release its models as open weight models in the first place? Like, many of the rest of the mob does this.

[1:15](https://realalpha.app/v/gJZLJSIEbtQ/?t=75) Well, there's some other Western labs that do it, but open weights has been part of the... Wait a bit. On and off. but open rates has been part of the core principles of Mistral since the very beginning when we started the sort of default was this technology which we really believe is going to transform everything, today it's transforming coding agents, tomorrow it's going to be finance, law all domains of science, material science biology, whatever comes to mind but this technology was going to be in the hands of a couple of established players and behind a closed API have the choice to turn the model off and decide who gets access to these to this intelligence and as we build more and more intelligent systems we think that actually this is it's not like the technologies of the last 20 30 years it's not like cloud infrastructure it's not like social media um it's not something which should be in the hands of very few really it should be democratized and for everyone to use and so this is where the open weights principle comes from it's a and it's been a guiding thesis for us from the very beginning.

[2:18](https://realalpha.app/v/gJZLJSIEbtQ/?t=138) Yeah, but you're also a business and you've got to make money. Yeah, so here I think the wins are in our favor. We had other players in the past that were closed-weight models, which left room for somebody to come like us, release a model to open weight, and build a business based on customization and fine-tuneability, which provides huge alpha also over a closed-weight model. So the idea here is that

[2:44](https://realalpha.app/v/gJZLJSIEbtQ/?t=164) **Sagar Vaze:** imagine you're an engineering company in Germany rather than being worried that the next generation of closed-weights models is going to come and wash over your competitive advantage.

[2:54](https://realalpha.app/v/gJZLJSIEbtQ/?t=174) Instead, the ideas you take in open-weights models, which in our vision is going to be exactly as strong as the closed-weights ones, exactly at the frontier, and you fine-tune it on your proprietary data, on your proprietary expertise. And then instead of your legacy expertise being something that you feel like you have to outgrow, instead it can be a competitive advantage in the AI age. AI becomes another tool for you to use, a competitive moat rather than something you have to be worried about. And this is something we find resonates very strongly with a lot of enterprises. Right. So on that point, Mistral has an outstanding niche when it comes to fine-tuned enterprise workflows.

[3:30](https://realalpha.app/v/gJZLJSIEbtQ/?t=210) Why is it better for an enterprise to use a fine-tuned Mistral model rather than just using Opus or just using Astra? um so different people have different perspectives on this um i think one argument which i've seen against against our thesis is that as these models become more and more capable um the benefit of fine-tuning is going to wash away and here this is actually not something which i've seen so even with our previous generation of models on and Actual real enterprise use cases, custom tasks, fine tuning still yields huge benefits today. So actually an example I like to give, it's a slightly more niche, but I think illustrates a point. Internally we have a very strong 30b model. A lot of flops went into it, which we use for sort of data annotation. This is a very simple task. Relatively we want to say, does this data belong to category A or category B? It's the kind of thing that you would imagine even models sort of two or three years ago being able to do out the box.

[4:31](https://realalpha.app/v/gJZLJSIEbtQ/?t=271) and a couple of months ago I remember running an experiment where we tried to do various prompt tuning strategies, some sort of harness around the model to boost our internal metrics. And ultimately we found that at the end,

[4:44](https://realalpha.app/v/gJZLJSIEbtQ/?t=284) **Sagar Vaze:** doing a small, what we call lower fine tuning, or a small fine tuning of the model on the data which we'd collected for this exact task was way, way better.

[4:52](https://realalpha.app/v/gJZLJSIEbtQ/?t=292) So it's a very simple task. Okay, sorry, it's true that the model is smaller than the frontier model of today, but it's also a very, very simple task. It's a simple classification data annotation task. And in practice, enterprise workflows are much, much more complicated. On things like, for example, specific coding languages which might be internal to a company, we're finding that even our Mistral Small from, I think it was Jan 2025, fine-tuned on these specific languages can now perform some of the best closed models. So in this space, it's always difficult to make predictions many years out.

[5:28](https://realalpha.app/v/gJZLJSIEbtQ/?t=328) But I see the advantage of fine-tuning and customizing lasting for a long time. Right. So with this niche, Mistral does very well with Enterprise, does very well with Sovereign AI. It is not, however, like a truly leading-edge frontier lab in the way that only Anthropoc and OpenAI are right now. So Mistral was founded in 2023. What do you think it would have taken for it to get there then? And do you think it's possible or desirable to get there now? Yes and yes to both, absolutely, unequivocally. What would have taken for us to get to the frontier in 2023 was us to be founded in 2020 or 2019.

[6:12](https://realalpha.app/v/gJZLJSIEbtQ/?t=372) I think this is really the key difference between ourselves, a very new lab, and some of the more established players, which is we just started a few years later. So we recognize that these labs are very capable models. the models that we've released today are actually state of the art in a bunch of key capabilities which are important to us finance, law, engineering some multimodal tasks close to my heart but even on the general agentic capabilities agentic coding where the best OSS model pushing the OSS frontier when it comes to saying what does it take to truly being catching up with OpenAI Anthropic this is absolutely on the cards That's absolutely the vision of Mistral.

[6:54](https://realalpha.app/v/gJZLJSIEbtQ/?t=414) This open weights category and closed weights category, it's a category that we seek to become closer and closer together. We don't want these things to have some delta in performance in the long term. And we think the thing that we need to get there is compute. So we've so far, I think, to date raised 6 billion, I forget, dollars or euros, one of the two, which is, I mean, a lot of money by traditional standards, But compared to what these other labs have raised, it's a drop in the bucket. And what we've done is build confidence in ourselves, in our technical chops, in our scientific understanding, that with this amount of capital, we can make extremely capable models.

[7:36](https://realalpha.app/v/gJZLJSIEbtQ/?t=456) As I said, frontier in many capabilities today.

[7:39](https://realalpha.app/v/gJZLJSIEbtQ/?t=459) **Sagar Vaze:** And so I think we've convinced ourselves that we can do more with less.

[7:43](https://realalpha.app/v/gJZLJSIEbtQ/?t=463) And I think the next step is to do much more with much more. so switching gears to the new model first of all where did the whole like le chaton fat and le chonk thing come from in the first place um what's like the story behind that okay you're gonna reveal me as a there's a bit of a boomer um so okay as far as i understand in the summer when some very powerful models were released um there was some sort of there was a post on on x um about mistral releasing le chaton fat um now as far as i understand there's actually a well-established meme history of different sized cats and le chaton fat is sort of the the biggest one um that's the one end of the spectrum of the meme so this this got a bit of traction and the funny thing was that at the time we were indeed training mr large four and so when we released it we called it le chunk and again as i said you're gonna have to forgive my like uh cursory knowledge of the meme domain but le chunk is sort of one step short of the chat on fat so it's a big model it's the biggest one we've trained but there's a there's bigger things in the works but le chat on fat was just invented by like the the x community it wasn't like an internal mistral thing uh no i think that

[9:02](https://realalpha.app/v/gJZLJSIEbtQ/?t=542) there was someone on x that invented it i think he's now got mistral stash i'm not sure how else we thanked him but yeah it was it was something on x which uh was funny because at the time we were all i think it was a joke because we not released a really big model in a while but it was funny because at the time we were creating our big model and the loss was going down nicely I want to shout out our sponsor, Lovable. You know, the app you've been meaning to build, the internal tool, side project, or product you'd otherwise lose a weekend to? With Lovable, you get software you can ship today. That includes an editable code base you can inspect and change, plus two-way GitHub sync.

[9:38](https://realalpha.app/v/gJZLJSIEbtQ/?t=578) Lovable also handles the backend and infrastructure, including managed Postgres, auth, storage, hosting, payments, and more, so you don't have to wire it all up yourself. And through Lovable's MCP server, you can create and deploy your project directly from the agents you already use. Turn ideas into software people love with Lovable. Lovable.dev. Now, on to the episode. So tell us about the model. What's new? What's different? What's exciting about Mr. All-Large 4? Yeah, so ML4, LeChonk, it's our first flagship OSS release in a while.

[10:14](https://realalpha.app/v/gJZLJSIEbtQ/?t=614) **Sagar Vaze:** It's a trillion parameter model trained on 1E25 flops, which is a measure of the amount of compute that went into it.

[10:20](https://realalpha.app/v/gJZLJSIEbtQ/?t=620) This sort of flop budget and parameter scale is sort of a key milestone in any AI lab's development. There was a bunch of infrastructure work that went into this, a bunch of things in the making. So first of all, actual real infrastructure. So we had to build out our own data centers. This model was trained from scratch on our own compute, which we built just outside of Paris. Also in pre-trained science, in order to be able to scale to the trillion parameter scale, being able to build predictors to understand the behavior of the models as you scale to this level and then also in post-training so to turn this strong pre-trained base into agentic capabilities requires a bunch of iteration on post-training we also did a bunch of research on that for the last 6 to 12 months and the result of it is this it's a ML4 LeChonk it's by far the best agentic coding model we've released both in our internal testing and with third party evals it's the best OSS coding model.

[11:21](https://realalpha.app/v/gJZLJSIEbtQ/?t=681) I think on the AA coding index, it's exactly tied with GLM 5.3, but needs 50% less output tokens. And then on these critical domains for us, which, as I said, are very important for our customers, things like cyber defense, finance, law, etc., where not only best OSS, but in many cases, the best model, period. Yeah, it's very nice to talk to open weights people because if I asked the same question to someone from OpenAI or Anthropic after they released their model, they would be like, oh, well, it's our best model yet. We're very excited. We put a lot of hard work into it, and that's all they would say.

[11:58](https://realalpha.app/v/gJZLJSIEbtQ/?t=718) So we are very excited. We did put a lot of hard work into it. But yeah, I think the magic of OSS is that you can share this with the community. We also benefit the other way around, right? So like the OSS landscape is in itself a community. In the past, for example, we would release, we released, I mean, people forget, we released the first open-weight mixture of experts model. So it was called Mixtral. I remember Mixtral back in the day. Okay, so not completely forgotten. But 8X7B was the first mixture of experts open-weight. and now a mixture of experts are sort of the underlying technology to every really large model today.

[12:34](https://realalpha.app/v/gJZLJSIEbtQ/?t=754) Then some of our friends in China, DeepSeek, they took this and they sort of turned the crank and they made this awesome V3 model, DeepSeek V3, which they made the model much bigger and they made it much sparser and then we in turn learned from that and if you look at our architecture that we released about a year ago, ML3, it bears a lot of similarity with the DeepSeek model. So that's sort of the magic of, one of the magics of open source or also the stuff that I discussed earlier, and we can discuss again. But as a scientist, the magic of OSS is that you can talk about your work, you can learn from others, and others can learn from you.

[13:09](https://realalpha.app/v/gJZLJSIEbtQ/?t=789) How is the model so token efficient? So this is something that we spent a bit of time on. Sorry, maybe I should clarify. We did a private beta release of this model through an API a few weeks ago. We used it to get early feedback with some trusted partners, and it's today that we're open-weighting the model such that anyone can download it. In that interim time, one of the things we worked on was token efficiency. So here, it comes through... I mean, we enforce this through reinforcement learning. So we'll release a full technical report, I think, tomorrow to discuss aspects of this. But we've got multiple reasoning modes and model from, I think it's low, medium, and high.

[13:51](https://realalpha.app/v/gJZLJSIEbtQ/?t=831) And the way it works is that we do RL for different checkpoints, which reason to different amounts, to different degrees, have a different amount of token efficiency and therefore different performance as well. And then we distill each of those into one single student such that one single model can reason in each of these formats in turn. So this is called a MOPD or on-policy distillation. Right. So another important area of model capabilities is cyber. And Mr. Alls talked about how ML4 is like the best open-weights defensive cyber model. It talks about how there's a very high refusal rate for malicious cyber requests, but also it's an open-weights model.

[14:34](https://realalpha.app/v/gJZLJSIEbtQ/?t=874) So people can download the weights, they can obliterate the model, they can make an offensive cyber tool with it. So do you see any tension there? Yeah, maybe to take a step back and again answer why we open-weights models in the first place, because I think this principle is broader rather than just cyber. Like I said, I think that this technology is going to transform every domain. And we fundamentally believe that this sort of technology shouldn't be in the hands of only a few established players, and it really should be accessed by everyone. So that's why we open the weights model in general, and I think the same principle extends to cyber capabilities.

[15:15](https://realalpha.app/v/gJZLJSIEbtQ/?t=915) So if you look at some of the most malicious attacks, or the most significant malicious attacks that have happened recently, I'm sure we can discuss many high-profile cases, But these have actually been perpetrated not by open-waste models, but by closed-waste models, which have either been jailbroken or, in some cases, during training, escaped their sandbox, etc. So it's actually, in practice, that's how many of the attacks have been perpetrated. And on the other hand, we've got defenders, enterprises trying to shore up their vulnerabilities, their own defenses. We're finding they're having their requests being refused by the same providers.

[15:53](https://realalpha.app/v/gJZLJSIEbtQ/?t=953) for completely understandable reasons, because in practice, these providers have classifiers under the hood, which say, is this a, is this a malicious? which is cyber request, or is this a regular request? And for the cutting-edge defense capabilities,

[16:08](https://realalpha.app/v/gJZLJSIEbtQ/?t=968) **Sagar Vaze:** these defensive requests in practice look quite similar to the offensive ones.

[16:12](https://realalpha.app/v/gJZLJSIEbtQ/?t=972) So it makes perfect sense that they are not precisely caught by the classifiers, that they're blocked by these providers. But the net effect of that is that people trying to shore up their defenses from malicious attacks, often from closed-waste models, are finding that they have no recourse but to use an open-waste model. And so we think that it's, I mean, we think that it's critical that there are open U8s models of the same strength, of sufficient strength, that people can really scalably defend their infrastructure. I agree, yeah. So another area is multimodality, which you work on. Tell us more about the multimodal capabilities of the model.

[16:48](https://realalpha.app/v/gJZLJSIEbtQ/?t=1008) What is that, where is it on the most important model capability now, which is being able to make really cool, like, 3D visualizations and models that you can post on Twitter? Yeah, so when we said that we wanted some early feedback, one of the things that indeed we found was that it wasn't as good at 3D generation as it should be. So I think you should cry out the new checkpoint, and I think it's pretty good at 3D asset generation. So I'm biased, but multimodality means all the modalities, and one of those is text. So it's a strict superset of all the capabilities. It's a strict superset of the capabilities, and in practice for us this means understanding images videos etc and on these domains we're finding that uh this model is really outstanding is it's the best uh it's by far the best multimodal model we've trained um it's also i think uh on all of internal evals the best oss model out there at the moment and on some capabilities uh for example visual grounding we're finding it to be the best uh uh you know competitive with all the best closed source also um for us this was a real technical journey so ml3 so our model from a year ago wasn't quite at the bar that we wanted and we made some some pretty big changes in

[18:00](https://realalpha.app/v/gJZLJSIEbtQ/?t=1080) pre-training uh pre-training and post-training to get to get the model up to scratch on on multimodal um particularly something i think we'll talk about in our tech report is that this

[18:09](https://realalpha.app/v/gJZLJSIEbtQ/?t=1089) **Sagar Vaze:** model is what we call a natively multimodal train from scratch so rather than training a model on text and then fine-tuning it on images and text or videos and text from the very beginning it's got images baked into it and so uh this has a bunch of interest infrastructure challenges

[18:24](https://realalpha.app/v/gJZLJSIEbtQ/?t=1104) because, you know, images are a heavier modality than text and also a bunch of scientific things to make sure that you don't regress on the critical capabilities like coding, et cetera. So a bunch of research to make sure that we could take multimodal to the heights that we think are possible and also to make sure that we didn't regress on text and that's sort of been the work for the last year on our team. Right, well, it's very exciting. Congrats on the model launch. Can't wait to use it. Thank you so much for coming on MTS. Thanks very much. Pleasure to be here. Cheers. A huge thanks to MTS sponsors.

[19:22](https://realalpha.app/v/gJZLJSIEbtQ/?t=1162) with serious security and governance. KongHQ.com. Finally prove what your AI investment is worth with Lairdyn. Head to Lairdyn.com today to start measuring your AI's bottom line. MongoDB, trusted to run critical applications by 75% of the Fortune 100, the intelligent data platform for the era. Apply today for access to up to $15,000 in Atlas credits, dedicated technical support, and exclusive events and go-to-market opportunities. MongoDB.com. Chargebee. Billing and monetization infrastructure for the AI era. Chargebee.com. Blitzy, autonomous software development for enterprise code bases. Ship 5x faster.

[20:04](https://realalpha.app/v/gJZLJSIEbtQ/?t=1204) Blitzy.com. AdQuick, make your brand a billboard. Out-of-home advertising as easy to scale as digital. AdQuick.com. Arena, measuring AI performance in the real world. Arena.ai. Support for the show comes from VCX, the public ticker for private tech. The U.S. stock market started history's greatest wave of wealth creation. From factory workers in Detroit to farmers in Omaha, anyone could own a piece of the great American companies. But today, our most innovative companies are staying private longer, which means everyday Americans are missing out. Until now. Introducing VCX, a public ticker for private tech.

[20:42](https://realalpha.app/v/gJZLJSIEbtQ/?t=1242) Visit getvcx.com for more info. That's getvcx.com. Carefully consider the investment material before investing, including objectives, risks, charges, and expenses. This and other information can be found in the Innovations Fund prospectus at getvcx.com. This is a paid sponsorship. You