AI NOW Summit 2026 — Paris — Opening Keynote

[00:00:00] the AI running your business wasn't made for you it was made for everyone you get a tool they get [00:00:10] control you get the same generic AI as your competitors no way to translate your expertise [00:00:17] into an edge no way to solve your unique business problems that's not power that's a limitation [00:00:24] your AI transformation starts with moving from artificial intelligence to actual intelligence [00:00:30] based on your business your knowledge and your work this is how you get AI that diagnosis an [00:00:37] issue on the factory floor in real time that helps banks deploy tailored solutions for risk and [00:00:43] compliance it's about custom intelligence whether it's running on a physical edge device for vehicles [00:00:49] or embedded in the complex operations of robotics innovators aren't interested in AI that was made [00:00:56] for everyone they're working to solve the world's hardest problems to lead their industry and to bring [00:01:02] ideas to reality that's why it's time to move from generic to tailored AI time to take control time to [00:01:13] own your AI transformation AI transformations accelerate here at the AI know summit where leaders reject [00:01:22] limits and chart their own path where the community drives initiatives that propel us all forward and [00:01:29] where innovators share insights from the past to build the future because history is made by those who act now [00:01:43] in the past to build the future and the past to build the future and the past to build the future and the past to build the future. [00:01:50] Hello everyone thank you for coming it's great pleasure for us to be hosting you in our first conference and so we will give a couple of updates on the things that we are announcing today it's going to be a very interesting day thank you for everyone for participating and thank you. [00:01:57] for all of the content provider and for all of the content provider and I'm here with Guillaume and Timothée and we're all going to present a video. [00:02:04] everything that we are doing on the on the various parts of the stack where we are operating on the various parts of the stack where we are operating so it's going to be a great day. [00:02:11] Thank you. [00:02:34] Thank you. [00:02:36] So let's get started. [00:02:46] Let me find this out. [00:02:48] First of all, well again thank you. [00:02:51] This is our first conference. [00:02:53] I would like to thank all of the organizers and all of the content provider. [00:02:56] I'm thanking also all of our sponsors that are here and have booths that you can come and visit. [00:03:01] I would also like to thank the team that have been organizing this event for the last few months. [00:03:07] And of course all of the Mistral team that have been working on all of the updates that we are going to give today. [00:03:13] Nothing would have been possible without you. [00:03:15] A lot of them are actually looking at us through the YouTube live stream if I understand correctly. [00:03:20] So let's get started. [00:03:22] We have two convictions at Mistral and we are building on two pillars for our strategy. [00:03:29] The first is that in order to deploy AI in the enterprise you actually need as an AI provider to own the full stack. [00:03:38] What do we need by that? [00:03:39] It means that we need to go from the compute and the chip operations all the way to the creation of business applications for the enterprises that we work with. [00:03:46] That means having multiple parts of... [00:03:52] That means having certain primitives for assembling business applications. [00:03:57] That means models, compute and I'll come back on that in a few minutes. [00:04:01] The second part which is as important is that AI is an horizontal technology that becomes powerful only if you go deep into certain verticals. [00:04:10] And if you focus on the actual business use cases where the creation of value can be achieved. [00:04:16] And I'll give some elements of that and describe what we are doing in the various verticals that we are focusing on. [00:04:24] Before that I'll start with a testimony of one of four partners. [00:04:28] ASML gives the chip producers the software, the hardware and the service they need to deliver on the promise of AI. [00:04:34] The combination of speed and precision really requires very sophisticated models and that's an area where AI is a great catalyst for ASML. [00:04:41] And that's where the partnership with Mistral will help us in a pretty formidable way. [00:04:47] Our ASML machines run 24/7 at our customer fabs. [00:04:51] Field service engineers need to diagnose issues as quickly as possible when they occur. [00:04:56] So we took the expertise of our experienced ASML engineers and combined that with Mistral looking differently at data and how to approach a problem. [00:05:05] We were able to develop a solution that's 120 times faster with a similar accuracy as we have today. [00:05:12] In the software development cycle the core issue that we are trying to solve is to catch the issues as early on as possible. [00:05:20] So that we prevent those issues from popping up at the customer. [00:05:24] Together with Mistral we built a custom solution for AI agents that meet our security and data needs. [00:05:30] These AI agents act as an always on code reviewer so that they can flag risky code changes and overall enhance our quality. [00:05:38] We think AI really creates a new era. [00:05:41] It's highly disruptive we think it's disruptive in the marketplace and a good catalyst for the demand for our tools for the offerings that we bring to customers. [00:05:50] And that's where the partnership with Mistral is really going to help us drive our efficiency targets and journey for the years to come. [00:06:06] Thank you. [00:06:12] So although we have grown and have really become a global company, we're still quite French. [00:06:16] So we're announcing the plan before presenting it and we're taking a rather deductive approach which is I think also very French. [00:06:24] Very Bourbacchist for the ones who have been doing mathematics. [00:06:28] So what does it take to actually build value for enterprises with artificial intelligence? [00:06:33] It takes compute because the models we're creating actually needs to run on large, on servers that consume a lot of electricity. [00:06:41] And really we are in the business of transforming electrons into tokens and intelligence. [00:06:46] And so the physical footprint actually matters, the control of the infrastructure actually matters. [00:06:51] So we'll give and Timothy will give some elements on that. [00:06:54] It does matter because we're currently in a situation where the supply crunch is very real on that domain. [00:06:59] And we have been at work deploying capacity for our customers to secure the quality of service that our customers need. [00:07:06] Then of course it takes predictive models that can handle different kinds of modalities and that can actually do the thinking in the language space. [00:07:15] This is how we started. [00:07:16] We believe that the foundation of all of the AI applications that are going to get built in the next years and the foundation of the new AI driven economy is going to be built on open source models. [00:07:27] And we've been at work for the last three years to actually make it happen. [00:07:30] And so Guillaume will take a look at the things that we have been releasing recently and give a few elements on what's coming as well. [00:07:38] Again, open source allows you to customize solutions, allows you to deploy the systems where you need them to be deployed. [00:07:48] Or if you want to deploy them on the edge, if you want to deploy them on infrastructure that you fully control. [00:07:52] Perhaps more importantly, it allows you to train the models. [00:07:55] And as you go and start using your own quantitative data, start connecting your models to proprietary AIP, [00:08:01] that ability to train the systems and to make them completely understand your business is extremely important. [00:08:07] Models are not useful in themselves in that it's a bunch of weights. [00:08:11] And so you need to be able to serve them. [00:08:13] You now need, if you want to build true business applications, to combine the token generators, [00:08:18] so the models themselves running on GPUs, to a set of primitives that allows you to take actions, to observe what's going on, [00:08:27] to actually improve the agents themselves. [00:08:29] And so we're deploying and bringing to market new primitives on the cloud space that allows you to do that. [00:08:35] We're also bringing a new product that Timothy will announce that is there to serve the end user in the enterprise and the developers in the enterprise. [00:08:44] I'll come back then to show and give a few elements on what we've been doing in multiple verticals, [00:08:49] and we'll announce a few new collaborations after that. [00:08:53] So with that, I'm happy to give the floor to Timothée, who should be there. [00:08:58] He's been lagging behind a little bit. [00:09:01] Is there anyone who can present what we've been doing on the computer site? [00:09:05] Thank you. [00:09:06] Thank you, Arthur. [00:09:10] Arthur, thanks pretty much everyone. [00:09:13] I'd also like to extend a warm thank you to you all for joining. [00:09:17] I didn't expect three years ago when we founded Mistral to find myself on a catwalk. [00:09:22] I cannot guarantee I'll use it to its full extent, but I'll do my best. [00:09:26] I also didn't fully realize three years ago that this journey would take us into the infrastructure world. [00:09:35] And so I'll talk to you about compute. [00:09:38] So about a year ago, we decided to take control of some of our infrastructure. [00:09:46] Compute is at the basis of everything that we do. [00:09:49] It's what we use for training models. [00:09:51] It's also what we use for inferencing and delivering all of our services. [00:09:56] This is a picture of our data center in Briar-le-Châtel in the south of Paris. [00:10:02] It was built in collaboration with Eclarion. [00:10:05] And we've been training models on it since the beginning of the year. [00:10:09] It's a 40 megawatts facility, and it's been very interesting to see how we can transfer rigor, [00:10:18] which is one of our company values, down to the hardware layer. [00:10:24] So applying all of that precision to fixing compute trays and fixing fibers, [00:10:31] allowing us to reach the very best speeds possible on that hardware for training. [00:10:37] This capacity is just the start, and that site will be fully in operation by the end of the summer. [00:10:46] Now, we've also announced a new site in Borlaingue. [00:10:51] I'm sorry, I don't speak Swedish. [00:10:54] But this is a site that we will develop throughout 2027, and that will host the newest Verirubin generation. [00:11:02] One of the benefits for us of owning the hardware layer is also that it lets us be at the very bleeding edge of what infrastructure provides. [00:11:12] And so with Verirubin, we hope to be one of the first to deploy this technology and provide new advances in efficiency and speeds. [00:11:23] All of this hardware would not be possible to use without software. [00:11:29] So we often talk about all of the investments, all of the ground development, all of the impact. [00:11:36] But to make all of this possible, you need to take that bare metal hardware and make it available. [00:11:43] We announced the acquisition of Koyeb, a company that develops serverless solutions. [00:11:49] This is especially relevant in the agentic world, where our customers, our developers, need a solution to host their MCP connections. [00:11:59] They need a solution to host their sandboxes, and they also might need to get access to GPUs. [00:12:05] And so it's been going on for a few months now, but Koyeb is fully integrated into Mistral, and it's been a joy to collaborate. [00:12:12] Now, today I'd also like to announce a new data center of 10 megawatts of capacity, again in the south of Paris, [00:12:21] conveniently located near to the Brier-Lechatel data center. [00:12:26] This is a 10 megawatt facility that will be live much sooner, this summer, and it will be used for all of our inference capacity. [00:12:36] More and more, the compute world has been getting supply constrained, [00:12:41] and so one of the reasons we've been doing all of this and developing all of these data center capacities is to secure compute capacity not only for ourselves, but also for our customers. [00:12:54] And so through this data center, we'll be able to offer a fully verticalized offering of Mistral products that we control from the application layer down to the hardware. [00:13:08] And with this, I'll leave it to Guillaume to tell you the latest on our models. [00:13:13] Thank you, Timothy. [00:13:26] Hi, everyone. [00:13:27] I'm Guillaume. [00:13:28] Thank you for being here today. [00:13:29] So I'm going to talk a little bit about our models, what we have built so far, also like what's coming next. [00:13:36] So every time we train models, we have a lot of things we care a lot about. [00:13:41] One of them, for instance, is open source. [00:13:43] As Arthur mentioned before, we have been releasing the weights of pretty much all of our models so far. [00:13:48] It was something we have been doing since day one, since the release of our first model, Mistral 7B. [00:13:53] It's also something we are doing even before Mistral. [00:13:56] This mission is really important to us. [00:13:58] It's one of the missions of the company to give AI to everyone, put AI into the hands of everyone so that everybody can benefit from AI on how it's powerful. [00:14:08] One nice thing you can do if you release the weights of your model, if you open source your model, is that you can customize them. [00:14:16] So you can take your model, you can train it on all your company data. [00:14:20] So sometimes if you have a company that is several years old or potentially that has been there for decades, [00:14:25] you have a tremendous amount of information that you have been collecting for years. [00:14:29] And what people don't realize before they do it is how much better a model can be if you fine-tune it, [00:14:35] if you basically continue the pre-training of this model on the data of your company. [00:14:39] It makes a very big difference compared to an off-the-shelf closed model that is going to be the same model used by millions of people. [00:14:46] The issue if you use a closed model is that every time you start a new conversation with it, [00:14:51] every time you start a new workflow, this new model will know nothing about your company. [00:14:55] So you will have to give it all the context every single time. [00:14:58] So it's going to work, these models are smart, but it's going to be very inefficient. [00:15:02] Well, if you train your model on all your company's data, in a sense the model will be trained on potentially millions of documents, [00:15:08] and it's going to know your company better than literally anyone in the company. [00:15:12] These models can encompass, they can integrate a huge amount of information. [00:15:16] So they will know the company very well, be very efficient. [00:15:19] So that's really something that we encourage people to do, that we do with many of our customers, [00:15:23] because there are massive gains when you actually do this customization on fine-tuning. [00:15:27] Something related to this is also having small and efficient models. [00:15:31] So since the very beginning, we have worked on compressing as much as possible, [00:15:35] the intelligence into our models, so trying to get a model as small as possible, but as good as possible. [00:15:40] When we released Mixtral 7B, it was the best open-source model of this size, better than much larger models. [00:15:45] When we released Mixtral 8x7B, it was like two years and a half ago, it was the best open-source model, [00:15:51] much better than much larger models as well. [00:15:53] So making small and efficient models has always been a focus for us. [00:15:57] And it was not that important like a couple of years ago. [00:16:01] So if you remember, before people were saying that models don't need to be this fast, [00:16:05] because people were only using this model for chatbots, or basically this multi-turn instruction following conversation. [00:16:12] But today the situation is very different. [00:16:14] Today we are building these agentic workflows, these models are running in the background, [00:16:19] they are doing a lot of actions, a lot of tool calls, so they are extremely token-hungry, [00:16:23] much more than before. [00:16:24] So what we are seeing today is actually a comeback of this small model and efficient model. [00:16:28] It was not something people were interested in before, but now it's very different [00:16:31] because of these new workflows that we are seeing today on these agentic capabilities. [00:16:37] Another thing that we really care a lot about is the multilingual capabilities. [00:16:40] So usually what we have seen is that people usually train models on the languages they speak, [00:16:45] the languages in the countries they come from. [00:16:47] For instance, the US model would be very English centric. [00:16:50] The one in China trained on English and Chinese. [00:16:52] For us, we really try to train our model on as many languages as possible since the very beginning. [00:16:57] The first Mistral model we are trained on dozens of languages. [00:17:00] The next ones that we are going to release are trained on more than 200 languages. [00:17:04] Basically every language our classifier can identify. [00:17:08] And this is really important. [00:17:10] It makes a big difference for many of our customers. [00:17:13] And it's also cheaper if you actually take this into account from the very beginning [00:17:17] because if you are dealing with a new language, a language that the model has not been trained on, [00:17:21] not only it's not going to be as good, but also it's going to be very inefficient. [00:17:25] Your tokenizer is not going to be used to this language. [00:17:27] So you need a lot of token to process a very small amount of text. [00:17:30] So yes, we have always put a lot of energy to basically make sure our models are really good [00:17:35] in all European languages, Asian languages, et cetera. [00:17:39] So the first model we released were text-only models, so fairly simple. [00:17:44] Models like a couple of years ago were much simpler compared to what we have today. [00:17:47] There was no agentic workflows, there was no multimodal. [00:17:50] So the first modality we added was basically giving the models the ability to process images, [00:17:55] which is very important when you want to process research papers, academic documents, [00:18:00] industrial papers with like a lot of figures, complex diagrams, et cetera. [00:18:05] Or even if you want to have your model being able to process slides and read them, [00:18:09] sometimes you cannot linearize into text the information. [00:18:11] So you really want the model to be able to read this information. [00:18:15] So now all our models are natively multimodal. [00:18:19] An example of a model that is multimodal and very small and efficient is Mistral OCR. [00:18:23] So it's a model that is very simple. [00:18:25] It does only one thing. [00:18:26] It takes some basically document and will convert it into a list of text and image [00:18:31] that you can feed into your large models. [00:18:33] So when we actually do this fine-tuning of models on top of our customers' documents, [00:18:39] when they have millions of PDF files or Word documents they want to train the model on, [00:18:44] the first thing we have to do is convert all of this into a format the model can read. [00:18:48] And this is where we use Mistral OCR. [00:18:50] So it's a very small model, one billion parameters. [00:18:53] It can process thousands of pages per minute on a single GPU. [00:18:56] And here if you work with some competitors, what they will recommend if you want to do this, [00:19:00] is to actually use a large model, ask it can you transcribe this document. [00:19:05] And it's going to work. [00:19:07] It's not a difficult task, but it's going to be extremely inefficient. [00:19:10] You will be using a model that is composed of hundreds of billions of parameters. [00:19:14] Here it's a 1B model, so extremely small, extremely fast, and it makes a big difference. [00:19:20] Another example of modality we added on the small model as well is Voxtral. [00:19:24] So it was our first audio model. [00:19:26] So we released the first version last year. [00:19:28] So initially it was just an ASR model that can do automatic speech recognition. [00:19:33] So it was basically transcribing. [00:19:35] So it was a 24B model, really good. [00:19:37] And a few months after this we released a model that is a 3B model, so 10 times smaller. [00:19:41] But also that can actually process more languages. [00:19:45] We also added a lot of features important for many of our customers, [00:19:48] such as language identification, speaker diarization, [00:19:52] so you can know who is saying what at which moment. [00:19:55] And yes. [00:19:56] And we also added some real-time capability. [00:19:58] So now you don't have to wait for the end of a speech to be able to process it. [00:20:02] You can process it on the fly. [00:20:04] The transcription will appear as you speak, [00:20:06] which is also useful if you want to do some live translation, for instance. [00:20:10] And a few months ago we released our last speech model, Voxtral TTS, [00:20:14] which is a model that can generate speech. [00:20:16] So it takes text as input, generates the voice. [00:20:19] It can take the voice that you want. [00:20:20] You can even clone your voice. [00:20:22] It's really efficient. [00:20:23] And now that we have done all of these different speech in, speech out on real-time. [00:20:28] Now we are combining all of these capabilities into a single model that we are going to release in a couple of months. [00:20:34] It's going to be our final Voxtral that will be a duplex model. [00:20:38] A model you can talk to. [00:20:39] It will talk to you back in real-time. [00:20:42] And, yes. [00:20:44] Here is another example of a model, Mistral Medium 3.5, that is a combination of many different models on capabilities. [00:20:52] So, I mentioned before we had like some multimodal models. [00:20:55] It was called Pyxtral. [00:20:56] So Pyxtral no longer exists. [00:20:58] It was a model that was really good at processing images. [00:21:00] But now all our models are natively multimodal. [00:21:03] So we don't have like a separate model for image. [00:21:05] It's going to be part of all of our models. [00:21:07] There is another model that we released one year ago called Magistral. [00:21:10] It was a model really good at doing reasoning. [00:21:13] So processing some complex math question or physics question, et cetera. [00:21:17] Now we no longer have Magistral. [00:21:18] This model is deprecated because all our models will natively be doing reasoning. [00:21:24] We also had some Devstral models. [00:21:26] That was a model for developers, really good at coding, writing code, writing PRs, finding issues, et cetera. [00:21:32] So here, the same, we also integrated these coding capabilities into Mistral Medium 3.5. [00:21:37] So we no longer have a Devstral model. [00:21:39] And the latest thing that we integrated is the agentic search capability, [00:21:43] which is basically giving the model the ability to do some search on some very complex environment. [00:21:50] So for instance, if you deploy a Mistral Medium into an enterprise, [00:21:53] and you ask it some question like, [00:21:55] "Can you summarize me this topic or tell me information about this or that?" [00:21:59] What the model can do is use some API calls to search the information on Google Drive, [00:22:04] on Slack, on Notion, wherever it might be, even potentially in your emails, [00:22:08] which is also why it's good if it's on premise or in your own server, [00:22:11] so you don't have to worry about confidentiality and privacy. [00:22:14] And this model will be really good at identifying if the information is there, [00:22:18] if not searching further, and sometimes searching like for maybe like 10 minutes, [00:22:23] like doing like dozens and dozens of tool calls. [00:22:25] And what we found is that by training the model to do this, [00:22:28] we reduced also a lot of the hallucinations, [00:22:30] because the model is trained to detect if it found the solution or not. [00:22:34] And if it didn't find the solution, it will tell you like that it doesn't have it basically. [00:22:38] So it's also reducing a lot of hallucinations, which is very important [00:22:41] if you want to deploy this model into some vertical and some industries. [00:22:46] So now we are continuing to add capabilities to Mistral, [00:22:50] to all our Mistral models. [00:22:51] So we are working now on the next large version of basically Mistral Large 4. [00:22:55] We'll potentially a priori be there in a couple of months at most during the summer. [00:22:59] We are adding many capabilities, for instance, in industrial applications, [00:23:03] such as like free dynamics, computational chemistry, computer-aided design, [00:23:10] but also like some cyber defense, cyber attack kind of capabilities, even finance. [00:23:15] Yes, so a lot of things that are coming soon that we'll be able to talk about more in a couple of months. [00:23:21] And now I'm going to give the floor back to Timothée, [00:23:24] who can tell you a bit more about the products powered by our models. [00:23:36] Thank you, Guillaume. [00:23:39] So this cat has been the playful face of a lot of our consumer-facing products. [00:23:47] It was first the face of our conversational assistant, Le Chat. [00:23:52] It was then the face of the Vibe CLI, so it's hanging around in your command line interface. [00:24:00] And as Guillaume mentioned, with the evolution of our models, with Medium 3.5, we understand the power of agentic models better. [00:24:12] Especially the power of manipulating the file system, manipulating a lot of information, and searching for what isn't there. [00:24:21] As we were using Vibe CLI more and more, connecting it to more tools, letting it do longer tasks, [00:24:27] we realized that this really didn't need to be bound to the CLI, it didn't need to be limited to code, [00:24:34] and we could do a lot more with it. [00:24:37] And so we are transitioning Le Chat to the Vibe family. [00:24:43] I have a short presentation video that will let you know everything that it can do for you. [00:24:48] I have a short presentation video that will be available for you. [00:24:50] I have a short presentation video that will be available for you. [00:24:52] I have a short presentation video that will be available for you. [00:24:54] I have a short presentation video that will be available for you. [00:24:56] I have a short presentation video that will be available for you. [00:24:58] I have a short presentation video that will be available for you. [00:25:00] I have a short presentation video that will be available for you. [00:25:02] I have a short presentation video that will be available for you. [00:25:04] I have a short presentation video that will be available for you. [00:25:06] I have a short presentation video that will be available for you. [00:25:08] I have a short presentation video that will be available for you. [00:25:10] I have a short presentation video that will be available for you. [00:25:12] I have a short presentation video that will be available for you. [00:25:14] I have a short presentation video that will be available for you. [00:25:16] I have a short presentation video that will be available for you. [00:25:18] I have a short presentation video that will be available for you. [00:25:20] I have a short presentation video that will be available for you. [00:25:22] I have a short presentation video that will be available for you. [00:25:24] I have a short presentation video that will be available for you. [00:25:26] I have a short presentation video that will be available for you. [00:25:28] I have a short presentation video that will be available for you. [00:25:30] I have a short presentation video that will be available for you. [00:25:32] I have a short presentation video that will be available for you. [00:25:34] I have a short presentation video that will be available for you. [00:25:36] I have a short presentation video that will be available for you. [00:25:38] I have a short presentation video that will be available for you. [00:25:40] I have a short presentation video that will be available for you. [00:25:42] I have a short presentation video that will be available for you. [00:25:44] I have a short presentation video that will be available for you. [00:25:46] I have a short presentation video that will be available for you. [00:25:48] I have a short presentation video that will be available for you. [00:25:50] I have a short presentation video that will be available for you. [00:25:52] I have a short presentation video that will be available for you. [00:25:54] I have a short presentation video that will be available for you. [00:25:56] I have a short presentation video that will be available for you. [00:25:58] I have a short presentation video that will be available for you. [00:26:00] I have a short presentation video that will be available for you. [00:26:02] I have a short presentation video that will be available for you. [00:26:04] I have a short presentation video that will be available for you. [00:26:06] I have a short presentation video that will be available for you. [00:26:08] I have a short presentation video that will be available for you. [00:26:10] I have a short presentation video that will be available for you. [00:26:12] I have a short presentation video that will be available for you. [00:26:14] I have a short presentation video that will be available for you. [00:26:16] I have a short presentation video that will be available for you. [00:26:18] I have a short presentation video that will be available for you. [00:26:20] I have a short presentation video that will be available for you. [00:26:22] I have a short presentation video that will be available for you. [00:26:24] I have a short presentation video that will be available for you. [00:26:26] I have a short presentation video that will be available for you. [00:26:28] I have a short presentation video that will be available for you. [00:26:30] I have a short presentation video that will be available for you. [00:26:32] I have a short presentation video that will be available for you. [00:26:34] I have a short presentation video that will be available for you. [00:26:36] I have a short presentation video that will be available for you. [00:26:38] I have a short presentation video that will be available for you. [00:26:40] I have a short presentation video that will be available for you. [00:26:42] I have a short presentation video that will be available for you. [00:26:44] I have a short presentation video that will be available for you. [00:26:46] I have a short presentation video that will be available for you. [00:26:48] Thank you, and thank you again for the team working on this. [00:27:00] It is beautiful this time of year in Paris, but setting a deadline for the end of May in France is vicious. [00:27:06] So there was a lot in this video. [00:27:12] It started with Vibe for Work. [00:27:14] So Vibe for Work is a new mode in our web app. [00:27:20] And I encourage you to go try it. [00:27:23] There is a guide that will take you through everything that you can do with it. [00:27:27] It will let you set up all of your connections. [00:27:29] It will guide you through its capabilities as well. [00:27:32] You can schedule tasks like, for me personally, what I would like is something that tells me what's going to happen in my week, [00:27:39] sometimes in my day, sometimes in the next hour. [00:27:42] You can summarize email and get this done every morning, every night. [00:27:47] So it's useful and I encourage you to try what it can do for you. [00:27:51] What was also shown was that it really connects to everything that you've built on Studio and for your company. [00:28:00] So it connects to your workflows that can process very complex and business logic tasks. [00:28:08] It connects to all of your managed connections as well. [00:28:12] It's really powerful that this tool talks to the same agent than the one you would use for code. [00:28:20] It's important to me personally because I sometimes code and less and less, but I also do other things. [00:28:26] And even when I did code, I was also interested in what was going on in my team and in the company. [00:28:32] So the agent that's in use for Viper work is truly the same than the one for code. [00:28:40] So when you access it through our web app or through the CLI, you have access to the same connections, the same tools, the same understanding of who you are, what you do, and what you're trying to achieve. [00:28:51] Viper code is also growing in the surfaces that you can use it on. [00:29:01] So it will also be available on the web. [00:29:04] With the growing power of these coding assistants, it becomes important to not be bound by a single CLI. [00:29:11] You often will see developers nowadays running one, five, ten sessions in parallel. [00:29:19] They might do this off of their laptops. [00:29:21] They might do this on the train. [00:29:23] They might do this on their mobile phones. [00:29:26] And so Viper code web lets you monitor all of those sessions on the go and basically fix bugs anywhere you are or add new features. [00:29:36] We're also including this with a new VS code extension that will let you again talk to the same harness with the same connections and all of the same tools available. [00:29:48] All of this is built on Studio. [00:29:51] Since the beginning of Mistral, we wanted to collaborate with large enterprises because we wanted AI to do things in the real world. [00:30:00] What I've quickly learned is that collaboration in an enterprise is a complex business. [00:30:06] We collaborate with enterprises with hundreds of thousands of employees. [00:30:10] We collaborate with governments. [00:30:12] In these environments, collaboration also means control and governance. [00:30:16] Studio is the place that holds all of the artifacts, all of the connections, all of the things specific to your business that the Viper work and Viper work and Viper code can access. [00:30:34] Studio is something that lets you govern what connections are available, what connections some teams can use and the ones that they can't. [00:30:45] It's also something that can be deployed on prem or wherever your data lives. [00:30:52] You can then segment the inference part from the data part. [00:30:57] So it really gives you control over the entire AI stack. [00:31:02] With Mistral Compute, you can even choose to run that inference on our own hardware. [00:31:11] After doing all of the work of setting up everything that you needed, the connections, the workflows, and really getting to know what Viper work can do for you, you will see that the models understand more and more about your company. [00:31:27] However, with general models, they will maybe not understand your company-specific DSL that has never seen the light of day. [00:31:38] They might not understand your 20-year-old code base that, again, has never been seen on the web or it might be in some esoteric language. [00:31:46] It might also not understand some infrared pictures of some mechanical components that you analyze for some reason. [00:31:56] To go that extra mile, we provide Forge. Forge is a tool that we deploy for our customers, usually where their data lives, to really do that last mile of adaptation between our model and the customer data and what matters to them. [00:32:13] It allows us to adapt to new modalities. It allows us to boost the understanding of our models for your data. [00:32:25] As Guillaume said, it also allows some optimizations. Once you've found that some tasks could be automated through Vibe, you will run it more and more. [00:32:35] Maybe it's something that the entire company adopts and the costs are going to rise or maybe you just don't have the infrastructure to run it. [00:32:42] Maybe that task is simple. You can take all of those traces, run them through a training on Forge, get a much smaller model to do it and save some dollars and maybe in the same time also accelerate that task. [00:32:56] This is the full offering of Mistral and the goal is really to give you full control and full customization capability around your AI because this is how we achieve true ROI. [00:33:08] With this, I'll give it to Arthur to explain how we do this with our clients. [00:33:17] Thank you, Timothée. [00:33:22] Great, so we are coming to our last part where I will discuss a few things that we've been doing in different verticals. [00:33:31] And I will start with one which is really exciting us quite a lot, which is Mistral for Industrial Engineering. [00:33:39] We've been working for the last year with companies whose core businesses and most important use cases of artificial intelligence are actually located in the R&D businesses in the business of creating physical objects. [00:33:54] And if you look at AI today and if you look at our announcements that we've just done, AI is great today at automating tasks for knowledge workers and for people that are doing software engineering. [00:34:04] So that's great because the software engineers can work faster. [00:34:08] That's bringing already huge productivity improvements. [00:34:10] But once you move to all the kind of engineers, well, they are underserved. [00:34:14] And why is that? [00:34:15] It's because the loop that you typically need to create in between the user and the test environment, which is the one you need to build in order to deploy AI systems on it, is much harder to control. [00:34:31] When you're doing, when you're modeling the, when you're modeling the wing of a plane or when you're looking at the process of your factories, you typically need to run simulations. [00:34:43] That takes a lot of time. [00:34:44] And you need those simulations to go way faster if you want to actually delegate some tasks to AI systems. [00:34:50] And so for this reason, we have announced last week the acquisition of a company called Amy AI, whose core business is to train models that understand different kinds of physics and can actually model the way a system will evolve over time. [00:35:06] Thanks to AI models that are trained on simulators and then retrain on the actual data coming from the sensors and coming from the true physical systems. [00:35:17] Why is that interesting? [00:35:19] Because we now have both the language intelligence and the physical intelligence models. [00:35:25] And by combining them together, we are building loops, we are building delegation loops that allows us to create better tools, that allows us to create better objects that actually have an impact on the physical world. [00:35:38] We've been at work already working with a very famous company that you may have used to come here. [00:35:47] So we're very happy today to announce our partnership with Airbus. [00:35:50] So we are working with them across their, their free entities. [00:35:54] So Airbus Helicopter, Airbus Commercial, Airbus Defense and Space on multiple domains, but everything centered around the engineering side. [00:36:03] So operations of engineering, we've worked with them on technical data. [00:36:07] We are working with them on design of the new products that are coming from Airbus. [00:36:13] And we are working with them on industrialization and the deployment of models into the aircraft themselves. [00:36:22] So Guillaume Fourie will give more color to that, I think in a few hours. [00:36:27] So again, thank you Airbus for your trust. [00:36:30] Great things to be done there. [00:36:35] We're also delighted to announce our partnership with BMW. [00:36:41] Again, a very famous engineering company that you may have used to come here. [00:36:46] We are working with them on engineering and on car crash simulation so that we can actually accelerate that part of the R&D cycle, which is actually the one that takes a very large amount of time. [00:37:00] So we're focusing on connecting our language models and our physical models to make it possible and to reduce that development cycle. [00:37:08] Again, thank you for your questions. [00:37:16] Industrial engineering is going to be a big focus for us. [00:37:18] And that means we have solution teams that are completely specialized there. [00:37:21] It also means that from a product side, we're going to be bringing Amy's model into studio and we will go after the different scales of physics. [00:37:31] We'll go after the micro scales, so computational three dynamics, etc. [00:37:35] All the way to the molecular dynamics that we need to model for certain of our customers. [00:37:39] So at the end of the day, that combination of physical intelligence and the language intelligence that we've been working for the last three years is the one thing that is going to have, we believe, a huge impact on the world. [00:37:51] Now coming to another vertical that is a historical one for us and also a very important one, which is finance. [00:37:58] We've also built a solution team that is completely dedicated to that domain and we work with multiple banks across the world to actually make models that allows better growth, better decisions on trading floors, that allows better back office operations and better customer operations. [00:38:14] So we're working with HSBC, we're working with multiple hedge funds in the US where our ability to deploy on prem and to actually help them build alpha that will not be shared across the industry is actually quite meaningful. [00:38:31] And I would like to reserve a specific place for our first customer or historical first customer which is BNP Paribas that has prepared a video as well. [00:38:49] BNP Paribas is a leading European bank operating in 64 countries and we serve all types of clients from individual to small businesses, [00:39:00] corporates and large institutional clients. [00:39:04] Gen AI offers major opportunity for more complex automation, but also some challenges. [00:39:09] So there are three main reasons why we chose Mistral at the beginning. [00:39:13] First is because many use cases that are valuable to us involve sensitive data. [00:39:18] So it was very important to find a partner that is able to help us run the models on our infrastructure. [00:39:25] The second one is that the models that are created by Mistral are more energy savvy. [00:39:31] And third, it seems very logical to support a European champion. [00:39:36] So within banking, we have a very important process, which is called QIC. [00:39:43] It's a long process. [00:39:45] So with Mistral AI, we started to work on a specific QIC process for corporates in Belgium and we developed two agents. [00:39:52] We were able to go from 80% of files incomplete sent to the middle office to only 10%. [00:39:59] For the clients, the application takes much less time, so we go from weeks to days. [00:40:06] From day one, we saw value in deploying Mistral's smaller models on BNP Parabas premises in order to satisfy our strong security requirements. [00:40:16] Our collaboration moved from models to solutions, research, products, and of course, we still have the models deployed across BNP Parabas group. [00:40:25] BNP tech teams and data science teams work directly with Mistral AI applied scientists, which means that they share roadmaps. [00:40:33] They work on common goals together, and in this way, we're developing agentic components for our LLM at CIB platform, which has rolled out to 65,000 users, everybody at CIB. [00:40:46] Having recently renewed the partnership with Mistral AI, not only the model, but also the software and the expertise, I feel really confident that we can initiate many other promising and valuable innovative projects. [00:41:03] Thank you very much. [00:41:06] Thank you very much. [00:41:08] We started to collaborate in 2023 where we were 15 people, so that was, I think, really a leap of faith at the time. [00:41:15] And we're very grateful that we've been able to continue that relationship across the last three years. [00:41:22] One other thing that is very dear to us as well as our mission is really to bring Frontier AI into everyone's hands, [00:41:32] is to be able to address the citizens and to address the way governments are actually interacting with citizens. [00:41:37] And so we have a program that we have no role to multiple countries, including Luxembourg, Singapore. [00:41:43] You will see Slovakia. [00:41:47] The Slovakian minister here also presenting what we have been doing with them. [00:41:51] We're working with France as well, with Caisse du Dépôt. [00:41:53] We're working with Morocco. [00:41:54] We're working with Greece on bringing models to the citizens in multiple ways. [00:41:59] So building public services, for instance, with France Travail, where we're deploying agents to help people looking for jobs. [00:42:06] With Luxembourg, when we're deploying systems that allows people to better understand the law of Luxembourg and to answer their administrative questions. [00:42:15] We're working with governments to make sure that models actually understand the cultural nuances of their countries in that they should understand the law in order for them to be interacting with citizens. [00:42:27] And so those are things that we've been doing with Morocco, for instance, to build models that understand Darijan and Amazigh. [00:42:33] We've been working with the Singaporean government as well, and HTX in particular, to build models that understand the Southeast Asian languages very well. [00:42:42] So, again, we think that AI needs to be specialized and understand the cultural nuances. [00:42:47] It needs to speak languages as good as it speaks English. [00:42:50] And for this, we do need to collaborate with states. [00:42:53] We need to collaborate on the data. [00:42:55] We need to collaborate on the training. [00:42:56] And then we need to look at what kind of applications can be brought to citizens to show the value of artificial intelligence that sometimes is something that is a bit frightening for them. [00:43:05] So we're very grateful of the trust that multiple governments in the world have been giving us, and we'll be continuing to do that. [00:43:15] I will end by an application which is really linked to our focus on multilinguality, which has been very important for us at the beginning, [00:43:25] because as a company headquartered in Europe, we had to face a diversity of language that is pretty unique. [00:43:32] And for this reason, Amazon is now working with us to make sure that on every interaction that are beyond English with Alexa Plus, [00:43:41] the quality is well controlled and the interaction is high quality. [00:43:44] So we're also very grateful to Amazon to be working with us on that domain. [00:43:54] So with this, it all starts now. [00:43:56] We are very grateful for the time, your precious time that you're giving to us today. [00:44:01] We'll make it useful. [00:44:02] You'll be able to see some demos. [00:44:04] You'll be able to see what we've been doing with customers. [00:44:08] You'll see a lot of customers' testimonies. [00:44:10] You should be able to exchange with our team and also with everyone here about the most important use cases [00:44:16] and where do you see the AI transforming your company. [00:44:19] We'd love to help. [00:44:20] Again, thank you and have a great day. [00:44:23] Thank you. [00:44:24] Have a great day. [00:44:24] Thank you. [00:44:24] Have a great day.