New AI Model Beats GPT4o While Running on a Laptop – Mistral Small 3.1

A new AI model from Paris is challenging industry giants with fewer parameters yet impressive power. Meanwhile, Tencent's tool is turning text into three dimensional designs in seconds, and Roblox is building entire game worlds with simple prompts. The AI race isn't just about size anymore. It's about speed, creativity, and who can change the game the fastest. Alright, so there's a brand new open source AI model on the scene. It's called Mistral Small 3.1 and it's coming straight out of Paris. Now, Mistral AI might not have the same name recognition as, say, OpenAI or Google's DeepMind, but it's quickly establishing itself as Europe's top AI company. They were only founded in 2023 by former researchers from Google, DeepMind and Meta, and they've already raised over a billion dollars in capital, landing them at a valuation of around 6 billion. All right, so Mistral Small 3.1 has 24 billion parameters, which is small compared to some of the biggest models out there. Its developers say it can outperform or match larger systems such as Llama 3.3, 70 billion parameters, and Quen 32B. Despite having fewer parameters, it's also multimodal and its context window reaches 1,128,000 tokens. So it can handle large documents and maintain context without dropping earlier details. Now, the speed factor is a really big selling point here. Mistral AI claims that this model can process information at a rate of about 150 tokens per second, which is seriously fast. And because the model has fewer total parameters, it can run on what they're calling relatively modest infrastructure. For instance, they say it'll work efficiently on a single RTX 4090 GPU or even a Mac laptop with 32 gigabytes of RAM. If you're used to hearing how these top end models need entire server farms or crazy clusters of GPUs just to spit out a few paragraphs, then you know that's a major shift. Mistral prioritizes refining algorithms and training methods over building massive models, which supports the broader goal of reducing costs and and environmental impact. Throwing more and more computing resources at bigger models has started to worry a lot of people, both from a cost perspective and an environmental one. Interestingly, deepseek is apparently doing something similar, so we might see more of these lean and mean AI models coming out soon. Now, Mistral AI isn't new to the game of releasing open source models. Previously they rolled out Mistral Small 3 and that version has already become the foundation for a bunch of specialized reasoning models that other researchers created. The open source strategy seems to be working because if you allow the broader community to tinker with your model, you'll get tons of innovations you might never have come up with on your own. And Mistral's going even further with that open source philosophy. Mistral Small 3.1 is under the Apache 2.0 license, so developers can pretty much do whatever they like with it, commercially or otherwise, without having to worry about restricted usage terms. Of course there is a trade off. Open sourcing your models can make it trickier to monetize them. Companies like Google, OpenAI and others have very closed proprietary systems and that exclusivity can help them build powerful business models. But Mistral is looking at revenue streams like enterprise deployments, specialized services, domain specific versions, or maybe partnerships with big tech platforms. They also have a line of specialized AI products which includes their flagship model Mistral Large 2, a multimodal model called Pickstral, a code generation model called Codestrawl, and these interesting edge device models called Less Mistral. For more niche scenarios, they just launched something called Saba for Arabic Language and Culture plus Mistral OCR to convert PDFs into markdown files for easier processing by large language models. Now, from what we've seen in performance benchmarks, mistrall small 3.1 is right around 81% accuracy on the MMLU benchmark. That's the massive multitask language understanding test, putting it in line with some 70B per minute models. The speed, the 150 tokens per second rate is also a big bragging rate and apparently in side by side human evaluations. Like real people judging the outputs, they found that Mistral's new release competes strongly against bigger systems on code, math, general knowledge and instruction following tasks. They also compare themselves to GPT4O Manny, a smaller but still closed model, and claim that they're outperforming it. So how do you actually get your hands on Mistral Small 3.1? They've made it available through a variety of channels. Hugging face is the obvious one if you're into the open source ecosystem, but it's also on Mistrol's own platform, Vertex AI by Google Cloud, and they're planning to bring it to Nvidia's NIM Microservices soon, plus Microsoft's Azure AI Foundry. They're also working with Ollama Kaggle together AI Fireworks AI, IBM Watson X and even more services like Amazon SageMaker, Databricks, Snowflake and so on. Basically they want it to be available Everywhere you might want to run or fine tune a language model Mistral AI is positioning itself as Europe's answer to tech giants like OpenAI and Google by focusing on smaller, efficient AI models instead of massive systems that demand huge computing power. Their strategy aligns with Europe's push for digital sovereignty, with support from figures like President Macron, who publicly encouraged citizens to use Mistral's Chat assistant instead of chatgpt the cloud. The company has secured partnerships with Microsoft, IBM and even the French army, while steadily expanding its lineup with models like Mistral, Large2, Pixdral and Codistrol. Alright, the next big story comes from Tencent, which has rolled out this incredible model called Hunyuan 3D 2.0. Now, since November 2024, these models have been aimed at designers and game developers, enabling them to transform text and images into three dimensional visuals super quickly. The buzz here is that it's almost instantaneous, meaning you type or provide an image of what you want and bam, you've got a three dimensional version ready to roll. This isn't just about speed either. The quality is reportedly high and there's a strong focus on consistency, which is huge if you're a designer who's always worrying about details like texture, lighting and overall aesthetic. Tencent has been making moves in AI for a while, but Hunyan 3D 2.0 shows they're serious about stepping into the generative AI game, which has mostly been led by US companies. Maybe Tencent. Their new model is said to outperform some big industry names, joining a wave of innovation from Chinese tech firms like Deepseek and bytedance. Earlier this year, Tencent launched the high speed Henyuwen Turbo S and now bytedance is rolling out its V Omniverse model. With both companies pushing hard in three dimensional design, things could get interesting in the digital content space. But there's a bigger global context here where China is coming up fast on the US in AI development. We're talking about potentially accelerating technological progress on a global scale and maybe even redrawing lines when it comes to trade and economic policies. After all, advanced AI is a strategic asset, and if companies like Tencent keep demonstrating the ability to innovate in areas like three dimensional design tools, we could see a new balance of power emerge in the AI landscape. Meanwhile, from the US side of things, Roblox has been making waves with the open source launch of Cube3D, a foundational three dimensional model that aims to generate objects, environments and eventually entire scenes, all from a text based prompt. The first piece of this release is the Mesh Generation API, which is currently in beta. This means that developers on the Roblox platform can crank out three dimensional assets in a matter of seconds just by typing in a simple description like a vintage green couch with clean lines and velvet material, and the system will generate a couch that fits that description. It's not limited to furniture, obviously. What's really interesting is how Roblox built Cube3D on native three dimensional data rather than trying to reconstruct shapes from two dimensional images. Think about it like a large language model, but for shapes. It tokenizes the various components of a three dimensional object, then uses an autoregressive transformer to predict what the next shape token should be. Kind of like how a text based AI predicts the next word in a sentence. This method apparently allows for more accurate and scalable three dimensional creation, and it also paves the way for multimodal input in the future, including text, images, video, audio or even three dimensional data itself. Now they are not just focused on creating three dimensional objects, they're aiming for full scene generation too. With their AI tools, you can describe a scenario and the system will predict both the layout and the objects to fill it. They're also working on what they call 4D creation, where objects won't just sit still still, they'll interact with players and each other in real time, and it's giving users way more freedom to shape their own game worlds. Generative AI is changing three dimensional design faster than ever, and it's exciting to see where this could lead. If you've got thoughts on it, feel free to share in the comments. Don't forget to hit that like button and subscribe if you haven't already. Thanks for watching and I'll catch you in the next one.