Mistral 3 Large AI Models Are Here!

So, Mistro just released the Mistro 3 family, and to be honest, I wasn't expecting much. Open- source models haven't really kept up this year. Every big jump in capability has come from closed models like GPT 5. 1, Cloud 4. 5, or Gemini 3. But this update is interesting, not because it beats everything, but because Mistral is clearly trying to make open-source relevant again, not with one giant model, but with a whole ecosystem. small models for edge, a new sparse flagship, and everything under Apache 2. 0 with real tooling support. So, let me walk through what's actually useful here, and then I'll show you a quick demo so you can get a sense of how these models behave. So, with the launch, we

actually get three next generation of MRO models. And these models are broken down by 14 billion, 8 billion, and 3 billion in parameter. And MRO large 3, which is their most capable model to date according to them. It's a sparse mixture of experts trained with about 41 billion active and 675 billion total parameters. And all of these models are actually released under the Apache 2. 0 license and they're open source. So this is amazing because they're dropping a whole layout of models that are supposed to be really good. And if we look at the benchmarks at the bottom here, we can see that compared to the other open- source models, Mistro 3 is actually leading in a lot of categories. When it comes to multilingual abilities, Mistro 3 is at the top with 85. 5 versus Deepseek is at 84. 2 and Kimmy K2 is at 83. 5. When it comes to the GPQA Diamond benchmark, Mistl once again comes at the top with 43. 9 and Kimmy K2 is at 35. 6 6 and Deep Seek is just a little bit behind at 41. 9. So we can see that across all these benchmarks, Mistro is consistently scoring higher except for two benchmarks which is the simple Q& A which is a little bit less compared to Kimmy K2. And then when it comes to live code benchmark, it's actually the least compatible compared to DeepSeek 3. 1 and Kim K2. So, Mistrol is really banking on the multilingual abilities as well, the GPQA diamond benchmark and simple Q& A a little bit, but when it comes to rest of the things, we're still seeing that Deep Seek and Kimmy K2 are leading. Now, when we compare it to Instructions abilities, we can see that Mistro 3 is at 53% while Deepseek is at 47% of the time it loses. So, Mistro 3 comes on top majority of the times. And this is the same thing when it comes to Kim K2. Mistro wins 55% of the time and Kim K 2 wins about 45% of the time. And this is only on general prompts. However, the real power is in the ability of the model to handle multilingual prompts. Deepseek against Mistral loses about 57% of the time and Deepseek only wins about 43% of the time. Now when we compare that against Kimmy K2, Mistro is actually winning at 60% of the times and Kimmy K2 wins at 40% of the time. So when we are looking at these new models, it's clearly showing that Mistro is supposed to be the leader when it comes to multilingual prompts and a little bit of a leader when it comes to general prompts. So this is huge for Mistro. So why would you choose Mistro 3 in the first place? Well, they say that it has frontier performance, open access, and it achieves closed source level results with the transparency and control of open- source models. This is kind of the mission behind Mistl, they want to be open source, and they want to give you that transparency, but don't want to compromise on the results. Then, when it comes to multimodal and multilingual abilities, you can now build applications that understand text, images, and complex logic across 40 plus native languages. So, it's something that is open- source. This is key for them because they want a wide variety of audience to be able to use their models and create amazing things with them. It's also very scalable. You can go from 3 billion to 675 billion parameters and you can choose the model that fits your needs. The model also has amazing agentic and adaptability. So, you can use it for coding, creative collaboration, document analysis, or tool use workflows with precision. I'm

in Mistro's AI studio and over here you can actually play with the various models that they have released. So you can do mistro large, you can select the small, medium, the 14 billion, 3 billion, 8 billion parameter models. So I'm going to test out the large model right now and you can tell the model to integrate the tools it needs. So a code interpreter, image generation, web search and web search premium. You can also give it functions and you can also add instructions like what model behavior you want the model to have like tone tool usage response style and then you can chat with the model here. So I'm going to ask the model to create a web page layout for a futuristic sneaker company and make it look clean and professional but engaging at the same time. So let's see what it creates. Okay, so this is the image that it generated of the new futuristic web page. Obviously, I don't think this is good at all. Uh, number one, the words are not clear. Uh, this doesn't make sense at the top. And as you can see, it looks more like AI sloped than an actual image. So, when it comes to image generation and that part of the multimodal ability for the new model, it's not the best, I would say. There's definitely so many better models out there now, especially with Nano Banana Pro and like that. So, I wouldn't necessarily use this model for UI mockups or anything like that. I'm going to ask the model to do some basic logic and I want to see its reasoning ability. So, I'm going to ask it. A shop sells notebooks for $3 each or a bundle of 5 for 12. You need exactly 17 notebooks. What is the cheapest possible cost? Explain your reasoning step by step, but in the fewest words necessary. After giving the final answer, summarize the strategy in one sentence and keep the explanation tight and logical. So, the cheapest possible cost is $42. So we can see that that's correct. The reasoning is bundle deal. Five notebooks for $12. So $2. 40 per notebook. Maximize bundles buy three bundles. 15 notebooks for $36. And the remaining two notebooks buy individually at $3 each. So $6. So total you get $42. Makes sense. Do you have an alternative is suggesting that which is more expensive which is two bundles plus seven individual. And the strategy it says buy as many bundles as possible then cover the remainder with individual notebooks. So this makes sense. So its reasoning capabilities are not bad at all. So that is good. Able to explain all of that in a very simple and clear manner. Now MR really prides itself on its multilingual abilities compared to many open- source models. So I'm going to test this out by asking it explain the idea of technological disruption in three ways. First give a clear explanation in English three to four sentence. Now continue the explanation but switch into French mid-sentence without restarting the thought and finally summarize the entire concept in one sentence in Spanish. Keep the tone consistent across all languages and avoid filler. So these are pretty three common languages across the world. But it's supposed to understand what has done in the first part using English and then translate that part remaining in French and then summarize everything in Spanish. So explanation is in English is here makes sense and then it goes into French. Um my French isn't the best but it's talking about Netflix over here and how it disrupted the video industry and Uber as well. And then it's summarizing all of this in Spanish. So, if you're a viewer and you understand French or Spanish, let me know what you think about the model's ability to translate this halfway into French, and is it coherent with the English? And does the Spanish summary over here do a good justice to all the explanation up above? Let me know. I'd love to know what you

guys think. If you enjoy this video, this is what we do here. Fast, clear updates on the biggest moves in AI. If you want to stay ahead of everything happening in this space, make sure you're subscribed. And if you want the hands-on side, demos, tools, workflows, and everything developers can actually build, well, check out the world of AI. We also run a simple no noise newsletter that gives you the most important AI tools and updates in just a couple of minutes. Subscribe here. Follow World of AI. Join the newsletter.