Liquid AI CEO Ramin Hasani's Keynote with Cristiano Amon at Qualcomm's Snapdragon Summit 2026

Yes, I'm Ramin Hasani, co-founder and CEO of Liquid AI. Started this company 3 and a half years ago, spun out of MIT CSAIL. We have been decade-long researching how to maximize the amount of intelligence capabilities you can get into smallest unit of compute. This is devices. So, and we wanted to enable physical AI. Like, you know, we wanted to enable AI agents inside the physical AI. Flash forward, Liquid AI. We are a foundation model company. We are building very efficient models that can be the core of agentic behavior. At the technology of our company at the core is a hardware-in-the-loop approach to really design foundation models from scratch. We don't do post-hoc optimization. You can do those post-hoc optimization, but we care about what we do beforehand. So, we design the models to be inherently designed for the device world. The technology that comes out of it is called Liquid Foundation Models, LFMs. These LFMs are very popular. We open-weight them. There are about 1.4 million downloads on a weekly basis on Hugging Face. You can actually see that. The models are ranging from 200 million parameters to 24 billion parameters. This is kind of a fabric of layers that I can enable like all sort of devices from wearables to like cars, you know, or maybe even PCs like bigger compute. We built two sets of products at this company. Around the models, we built something that we call model plus X, model plus a harness that is ready to get deployed on top of device. This is product number one. And product number two is a model development stack that allows us to take these opportunities and give OEMs and in a B2B fashion, give OEMs the opportunity to really customize this model plus harness for their applications and downstream kind of use cases that they want to do. The goal is like basically is very ambitious, you know, we want to be deployed, we want to have LFMs deployed on 50% of words annually produced devices across wearables, handsets, automotive, PCs, robots, and industrial devices.

We are seeing that AIs are becoming very very capable. Capabilities increasing exponentially. We're seeing fundamental math innovations. We're seeing better reasoning models. We see longer and longer time and more complex workflows are getting kind of done. But there's always a disconnect, you know, so you always see that individual product, if you look at the metrics, individual productivity based on these AI models that we have today is way ahead of business transformation. The gap is always between how are you bringing these capabilities and matching them to value, to actual value. That gap has been the problem. So, I feel like on the AI journey, we have been on this, you know, the fundamental question that we have to address is production grade AI. How can we make AI to be useful? How can we like increase the amount of use of AI? That is something that is extremely difficult to get. On this journey as well, like what does production AI actually mean? For me, it means reliable actions, agents that can take reliable actions, quality of agents that are really high at the level of kind of frontier, you know, that we're seeing. Verifiable outputs, the outputs of the systems are completely verifiable. And at the same time, you are satisfying the deployment constraints that you have, you know, use cases, they might need privacy, they need connectivity, they need, you know, like many different aspects, security and sensitivity. So, and then there's use cases and then there's deployment environments, you know, you have the physical world deployments, you have the cloud deployments. So, you got to be able to satisfy those things also with the technology. And I think the next wave of this technology, basically hardware and kernel and model and agents, all basically co-design of these things together is something that I'm looking forward and I think we're on the right path together.

I think five things has to change. Okay, so the first one is quality of we're talking about device intelligence. We want we need to reduce the gap between frontier level performance on the cloud and devices. That is something that is like absolutely important. So, if you do not one of the reasons why device AI and local AI revolution hasn't happened yet, for me is just that quality bar. Getting to that quality bar is number one. Number two is context. When we're thinking about devices, we're talking about a fabric of devices. Users want to use AIs on one device and then be able to transfer their experiences and preferences to other devices as well. How do you do that? You need to have the context layer. You need to have something that allows you a memory layer that connects these devices together. Third thing I would say is a harness. Harness is all the toolings that goes around the foundation model. When we talk about agentic AI, we're talking about a model plus a harness. That harness, the adaptation of the harness, and the adaptation of a model should happen at the same time. The fourth thing I would say is understanding the limitations of the physical world. You know, so for example, we're talking about the real world has constraints, devices have constraints, understanding the constraints of processors, and then feeding them back to the model design process. So that becomes the co-design process. And the last thing I would say is always developers, developers, developers. We need basically an army of developers. We need to open source. We need to give developers access to the technology to build, to really bring the AIs to billions of devices. You know, so we're talking about local AI. That transformation needs attention from indie developers. And for enterprises, you know, data is the most important thing. I believe that model building capability, we should bring products into the hand of enterprises. So model building capability becomes an ownership from the enterprise side. So enterprises should own the model building capability. So you just give them the framework so that they can get enabled.

So the next wave, think about where we are today. We are building like you see massive amount of investment going into the cloud business, which is great because we have exponential amount of demand for AI. So there is a lot of investment is going into that direction. I think the next wave is going to be local AI.

Why? Because every year we have more than 2 billion processors getting built that can host a foundation model on top of around the world. Every year 2 billion devices. This is an untapped opportunity. It takes the qualities that I mentioned to really enable that local AI revolution. And off of these processors, I think Qualcomm is like best positioned because you own a whole lot of that population of the 2 billion. And I think that's the next hot wave.

Very excited to announce today a product that I'm really proud of, Liquid Context on Snapdragon. This is a smart memory layer that allows you to read, compress, and write every sort of experience that you have throughout your let's say active engagement with the device. And across devices, this context layer is going to enable AI systems to go from one device to another device. That is what we are announcing today. Systems that are running locally very very close to the user. This Liquid Context that we're announcing, it has two major components. One of them is the memory component that allows you to gather information and from all sources like abstracting away like let's say preferences of the users and building that memory layer for an AI agent or army of AI agents that are running locally on a Snapdragon kind of chip to really access those information and perform jobs. So, this whole these two products together, the memory layer and the agent layer, are coming together. Actually, so let me show you a demo about like to inspire you. This is a B2B product. So, OEMs are going to be using this. So, let me show you a product just to inspire like how does it look like. Here is actually my agent. This is running on a device. This is running on a Qualcomm Snapdragon. It's actually directly running on an NPU or the Hexagon. So, it comes with my permission it consumes all of whatever I give it access to. So, it's basically senses everything that I give it. My Slack messages, my notes, my photos, everything that I do in my life I give access to this. Why do I trust this agent? Because it runs on my device. I give it access. It actually enables agents that are running on the NPU and does proactive job for us. The innovation here is making agents to understand when to bring you value. This is proactiveness. This is not reactive. You're not asking the system to do something for you. You can. This is the old generations. We're talking about proactiveness. This system actually designed for us. You know, it manages my calendar and at the same time it also passes through, you know, like it drafts social media kind of message. It knows that I want to post something on X all the time and it actually drafts the messages for me, makes it ready for me to go. That's an example of what the OEMs can do as the final kind of product. But, the whole goal here is that we want to prepare this as a product together to really bring it to market to enable enterprises across cars, mobile phones, laptops, wearables to be able to access a shared context and very powerful agents running on a Snapdragon to really enable a full experience for users.

That's true. Yes, in Mercedes-Benz cars like we have a partnership with those guys. Like we are running completely local experience of enabling in-car intelligence and that's a reality that you make it real. And this is production grade AI, you know, deployment actual deployment of these things on Snapdragon chips inside cars.

The degree of sophistication of this use case itself, I'll tell you that drives you next to the device. You got to be on the device. You got to be having a powerful system that runs with the user. The user has to have always access to a system that is continuously monitoring and continuously capturing information. Snapdragon enables that. The other thing that I would tell you is the heterogeneous kind of design of the platform itself. We use all components of the processor, you know, like we're using the Hexagon for running the agents very efficiently. We're using the sensing hub for really abstracting away like bringing all the components that we can read basically like whatever device metrics that we can read. We use the CPU for agentic kind of sampling, you know, so there's like all sort of kind of elements of the Snapdragon that enables this. Which is really beautiful for us. And also, one more thing to tell you like when we started Liquid AI, one of the most popular CPUs in the market was Qualcomm CPUs, you know, so we actually designed the second generation of Liquid Foundation models like for Snapdragon. So, the models are actually truly optimized already for the Qualcomm chips.