Mira Murati's AI Model Inkling is Open-Source ... but Huge

Mira Murati's Thinking Machines open-sourced their AI model but don't expect to run it even on a high-end PC. Find out more in today's presentation. Today we're talking about Thinking Machines Lab and their highly anticipated new AI foundation model, Inkling. Now on the surface, this looks like a massive win for the open-source community. It's a wildly capable AI system that you can just go and download. But as we peel back the layers today, you're going to see that what looks like a generous gift to the public is actually a highly calculated multi-billion dollar business gamble. Let's get right into it. Okay, so let's dive into this massive contradiction sitting right at the center of the whole thing. Inkling is completely free to download. It's got a super permissive Apache 2.0 license, which basically means, in theory, it belongs to everyone. But in practice, well, it is completely, 100% impossible for a normal person to actually run this thing on their own hardware. It really forces us to ask, what good is a free AI model if you literally can't afford the electricity and the hardware to turn it on? To figure out why they built it this way, we need to look at exactly how open this model really is. And this perfectly illustrates a distinction we absolutely have to make when we're talking about modern AI. Inkling is an open-weights model. It is not a fully open-source project. Think of it like this. They are generously handing you a fully baked cake, right? You can eat it, you can slice it up, you can share it around. But they are entirely hiding the kitchen they used to bake it. You don't get the training dataset, you don't get the data selection pipeline, and you certainly don't get the pre-training recipe. You just get the extremely heavy final product. Now, the sheer scale we're talking about here is literally wild. Inkling is an absolute behemoth. We're talking 975 billion total parameters. But it's built on what's called a sparse architecture. What that means is, instead of using all 975 billion parameters to answer a simple question, it acts kind of like a smart manager. It efficiently routes information to specific expert clusters within the model, so it's only using about 41 billion active parameters per word it generates. That makes it computationally way faster to run than a dense model of the same size. But, and this is a massive catch, you still have to store all 975 billion of those parameters in your active memory. And that storage reality? It hits you like a freight train. Even if you use the absolute most compressed, quantized version of this model, meaning the bare minimum, squished down file, it still demands a staggering minimum of 600 gigabytes of aggregate GPU memory. And that is just to load the weights. That doesn't even account for the extra memory you need to actually process your prompts and generate the answers. Let's put this into perspective, so we can see just how impenetrable this wall is for everyday developers. Your absolute top-of-the-line, break-the-bank gaming PC? It might have, what, 24 to 32 gigabytes of VRAM. Trying to run Inkling on that would probably just melt your motherboard. To even boot up the compressed version, you need 600 gigs. And if you want to run the full-quality uncompressed version, you are looking at over 2 terabytes of GPU memory. That requires 16 top-tier NVIDIA H200 datacenter GPUs. So yeah, this isn't just out of reach for a hobbyist, it is way out of reach for most small businesses. Now you might be wondering, wait, didn't they release a smaller version? And yes, they did preview Inkling Small. But please do not let that name fool you. At 276 billion parameters, this so-called small model requires a minimum of 138 gigabytes just for the weights alone. So sure, it might be a bit cheaper for massive enterprise datacenters to run, but it is still strictly a heavyweight commercial model. It's completely impractical for an enthusiast's local desktop. The word small here is purely relative to datacenter standards. So that brings us to the obvious question. Why? Why would one of Silicon Valley's most heavily funded new AI labs spend immense resources and, frankly, immense amounts of compute to build a frontier-scale foundation model only to give it away to a public that literally cannot turn it on? Why give away a product that costs an absolute fortune to run? Well, to make sense of this, we've got to look closely at the people behind the company. To really understand the strategy here, we have to neutrally look at the founder and CEO Mira Murati. She was the chief technology officer at OpenAI, right? She led research and product during their most explosive period of growth. During that chaotic November 2023 weekend, you know, the OpenAI blip, where Sam Altman was temporarily ousted, she stepped up as interim CEO before ultimately supporting his return. By September 2024, she left OpenAI completely. And just a few months later, in February 2025, Thinking Machines officially emerged. This context is absolutely vital because it shows a leader who deeply understands the very frontier of AI, but also acutely understands the vulnerabilities of massive centralized labs. And listen, that kind of elite pedigree does not go unnoticed. In July 2025, Thinking Machines secured a truly extraordinary $2 billion seed round, all at a reported $12 billion valuation. That is staggering for a company that, at the time, had no public product and zero revenue. But when you look at the backers, we're talking heavyweights like Andreesen Horowitz, and critically, NVIDIA, it becomes super clear. They were funding a team that had already proven they could build world-changing AI at scale. So they took that massive cash injection and used it to aggressively recruit top talent, grabbing heavyweights straight out of OpenAI, like John Shulman and Lillian Wang. But it hasn't been entirely smooth sailing. The AI talent war right now is fiercely competitive. We've seen some pretty significant leadership instability. Key co-founders like Barrett Zoph and Luke Metz have actually already left and gone back to OpenAI, and others have reportedly jumped ship to Meta. Honestly, the fact that Thinking Machines still managed to ship a $975 billion parameter model amidst all this intense corporate drama, it's actually a massive testament to the technical bench they still have. And through all of this, they shipped inkling, wrapping the whole thing in a very specific philosophical stance. They argue that AI shouldn't just be controlled by a few centralized, closed-door laboratories pushing generic APIs. Instead, they claim AI should be deeply customizable. Their pitch is that organizations should be able to own the weights, train the models further on their own private data, and really shape these systems with their own judgment. It's a very compelling, democratic-sounding narrative. And honestly, it sounds fantastic on paper. But if we strip away the lofty philosophy for just a second, the real business strategy comes sharply into focus. Inkling is not a charity project. It is essentially a Trojan horse. It's a massive, shiny, attention-grabbing asset designed to do one thing — attract highly lucrative enterprise customers to their actual product. And that product is a paid, commercial platform called Tinker. Tinker is a managed API for model training, fine-tuning, and deployment. It's honestly a brilliant move. This is your classic razor-and-blade strategy, just updated for the AI age. Step 1. They give away the massive model weights completely for free. Step 2. By doing that, they completely eliminate a massive corporate fear, which is vendor lock-in. A huge bank or a hospital knows they legally own their customized model and can walk away whenever they want. But step 3? That's where the trap snaps shut. Because the model is impossibly huge, that bank or hospital simply cannot run it themselves. They absolutely must pay Thinking Machines for the Tinker infrastructure to actually train and operate the thing. You see, they aren't trying to sell the AI model itself. They're monetizing the incredibly scarce ecosystem surrounding it. They charge for fine-tuning usage, for hosted inference, checkpoint storage, and dedicated enterprise capacity. And how exactly do they provide all this massive compute? Through a 1-gigawatt data center partnership with NVIDIA. They are essentially saying, hey, the model is yours, but the computational power and the operational expertise required to actually use it? Yeah, you're gonna have to rent that from us. So we kinda need to redefine what open AI actually means in this specific context. Thinking Machines isn't trying to democratize AI for the average consumer. Or the indie developer. Or some student coding on a laptop. Not at all. They are democratizing what we should really call institutional model ownership. They are building the heavy-duty tools for massive enterprises, banks, and research labs who want their own specialized private AI brains, but who really don't want to build a billion-dollar distributed supercomputer from scratch just to get there. Which leaves us looking at the future of this incredible gamble. Thinking Machines prove they can build frontier-scale models for sure. But building a sustainable business around it? That is a whole different beast. Leaving us with one final provocative question to ponder today. Can the Tinker platform actually win over enough enterprise customers to justify its massive $2 billion funding? Or will the immense crushing weight of gigawatt infrastructure costs and ongoing executive turnover ultimately sink them before their vision is even realized? It is definitely going to be fascinating to watch.