CoreWeave Fully Connected 2026 | Jeffrey Li, Liquid AI

Clips
More from CoreWeave Fully Connected 2026

Jeffrey Li

COO

Liquid AI

play_circle_outline Liquid AI vision: run production-quality AI models on devices at the edge

play_circle_outline Liquid Context with Qualcomm: device-captured personal context layer for agents

play_circle_outline Lightweight Agent Harness and Efficient Context Compression for Constrained Devices

play_circle_outline Scaling Edge AI: Partnerships, Model Portfolios, On-Device Self-Healing, and Hardware Challenges from Phones to Servers

Info
Transcript

Jeffrey Li, Liquid AI

Jeffrey Li

COO Liquid AI

In this interview from Fully Connected 2026, Jeffrey Li, chief operating officer of Liquid AI Inc, joins theCUBE's John Furrier and Dave Vellante to discuss how Liquid AI is redesigning foundation models from the ground up to bring production-quality intelligence to edge devices. Li explains why traditional transformer architectures, with their exploding KV caches, cannot fit within the fixed compute of phones, cars and wearables, prompting Liquid to build non-transformer models suited to constrained environments. He details Liquid Context, a new layer develo... Read more

explore Keep Exploring

What is the company's vision for bringing AI to edge devices, and what product did it announce at the Qualcomm Snapdragon conference? add

How does your new "Liquid Context" work to capture personal context at the edge (on phones, wearables, cars, etc.), and why does that contextual information need to be gathered on-device rather than in the cloud? add

How must agent harnesses and models be re‑architected and optimized to run on edge devices with fixed compute and storage (e.g., handling user memory, context size, and long‑term storage), and what trade‑offs—such as avoiding transformer architectures—does that entail? add

What is the company's current focus, how is it working with CoreWeave, and what are the main challenges to scaling edge AI deployments? add

bolt Powered by CUBE AI

Jeffrey Li, Liquid AI

search

(INTRO)

John Furrier

>> Welcome back, everyone, to theCUBE's live coverage of CoreWeave's Fully Connected Conference. I'm John Furrier, with my co-host Dave Vellante. We're talking about AI infrastructure systems that are powering the next generation AI applications. The next guest is the CEO of Liquid AI, a company we've featured on theCUBE in our NYSE Wired brand. Jeffrey Li's here. Good to see you. Thanks for coming on. Your company's been on 3 times, you're the

>> 4th.Yes, thank you for having us. We've been really excited. I'm glad to represent Liquid for the first time

John Furrier

>> personally.Yeah, well, first of all, Love the firm, fan of the company mission, but a bunch of smart people spun out from MIT. You guys are working on a really interesting model. Explain the vision for the folks who haven't heard it before of Liquid AI.

>> Definitely. The vision we have is that we should bring AI, the same AI that everyone is familiar with in the cloud, we should bring it closer to the user. So we focus on building these AIs to run on the devices all around us. So in your phone, in your car, in your PC, and these AIs traditionally are extremely compute and resource hungry and need large data centers. And so we have to redesign them from the ground up to be able to fit and run at production quality in the devices all around us.

John Furrier

>> Our last conversation was about the edge. We love talking about the edge, but you guys have news. So you just came back from the Qualcomm Snapdragon conference in Hawaii. Share the news.

>> Definitely. Yes. So our origins have been about building these efficient models. And beyond simply building the models to perform agentic tasks, we also announced a new product we're calling Liquid Context. We're doing this in partnership with Qualcomm. What it does is it actually sits between every model and agent and the hardware, and it focuses on helping the device build a real understanding about you, who you are and what you're trying to get done. And it gives this information that only your device is positioned to capture to any agent so it can be more proactive and personal to you.

John Furrier

>> Talk about how that works. the iPhone is basically a supercomputer from the old days, but it's still limited compared to the AI factories that are pumping out intelligence. What's going on at the edge? Explain how it works.

>> Fantastic question. So why personal context has to be something that is captured at the edge comes down to the context about you that matters, that that the gap between the context your Manus agent or your Claude agent has today and what you really need is what's happening to you in the physical world. It's not just about your digital footprint. Who are you sitting with? How are you feeling? What are you trying to get done? Where have you been? These are the questions that really help a human assistant do things for you proactively. And all that's invisible to things that live and breathe in the cloud. What is the best form factor to capture that signal? It's the devices in our pockets. It's our phones, it's our wearables, it's our watches, it's our PCs, it's our cars. And so it's not even just that we can do it in the devices to build this context layer that helps agents be more proactive. We believe you have to build it in the devices, and Qualcomm is a partner that's in all the devices.

Dave Vellante

>> So let's double-click on that. One of our analysts, George Gilbert, likes to say there's more agent harnesses than there are fleas on a camel's back. And so what happens when you try to put one of those harnesses on a phone in a car? What are the challenges? And talk about your approach of from the ground up.

>> Yeah, fantastic question. So the harnesses, at the end of the day, a harness is simply the scaffolding that turns a model into an agent. It gives it memory about you. It allows it to call tools on your behalf. Now, what we have seen is the problem with these harnesses, just like the models, is they're built with the assumptions of the cloud, which is if you need more compute, you simply increase and scale your compute. The problem with devices is that you have fixed compute. You can't, if your agent or harness needs more compute than your phone provides, you can't just add more RAM or scale it. You have to fit within the zero-sum compute. And so that means a lot of the assumptions around how harnesses today are built no longer hold at the edge. And I'll give you one concrete example. So one of the things that makes a harness really great is if it understands you over time. Many harnesses, OpenClaw, Hermes Agent, what they do is they build an ever-growing text file of your context. And that's perfect. Perfect when you have unlimited storage.

Dave Vellante

>> Where's that live? Yeah, right.

>> Exactly. But at the edge, you can't do that for many reasons. You don't have infinite storage. And also the models cannot consume infinite context because of the limitations of the compute. So that means you have to re-architect the harness. You have to redivide the responsibility of understanding the user between the harness and the model and the context layer to be able to maintain that functionality of personalization within the constraints of the real-world device.

Dave Vellante

>> So you're saying with mainstream harnesses that would sit on an object store somewhere, cheap and deep storage, So can you explain a little bit more about how you optimize? Presumably you're optimizing for memory and longer-term storage.

>> Definitely, yes. So we optimize both the model and the harness layer. So on the model side, what we're optimizing for is we're looking at what are the main bottlenecks that cause traditional transformer-based models to not fit on devices. It's often to do with the large exploding KV cache of a transformer-based architecture. Our models don't use transformers. Then the way we have gone about this goes back to the origins of our company. We've always been in the pursuit of more efficient building blocks and operators to build models and understanding how we fit within the constraints of compute on the device but still bring production quality. And on the harness side, what we also look at is what are the key functionalities of an agent that people expect when it's running on a device and which one of those components, how they're traditionally built in the cloud, such as pushing to Object Store, no longer work. and how can we redesign those from scratch? So, for example, one of the things we don't do is we don't build an always growing and appending log of memories. We have our own models that are very thoughtful about when to store what, and also how to compress and store this in a format that is more efficient, as long as it's readable by the agent, even if it's not readable by the human.

Dave Vellante

>> So Transformers are hogs, but they're powerful. So what's the trade-off?

>> Yeah, fantastic. The Transformer is powerful, because it is able to look back at a lot of context and generalize very well. And so you see these transformer-based models, they are able to perform at very high quality across a range of tasks. For the models that we build that don't use transformers, we were able to be much more efficient. The trade-off is the general-purpose knowledge and capability of the models are slightly worse. But that's okay because at the edge, especially in the small regime, you're not trying to deploy a general-purpose competence. The analogy I like to use is the cloud is a PhD or beyond PhD at every subject. People at the edge are not looking to settle for a high schooler at every subject. They're looking for a PhD at what matters. And when you constrain to what matters, we are able to fit that narrow quality into the constraints of a non-transformer.

Dave Vellante

>> So purpose-built for a specific task, if you will.

>> Yes.

Dave Vellante

>> And that's, presumably, a very large number of tasks that you can optimize for.

>> Exactly.

John Furrier

>> All right. So when agents cross boundaries, that's a big thing we're talking about with sovereignty. But let's just take it for a second for you guys. When an agent crosses cloud to device, what breaks? What problems does that solve? Because agents need to go to the device. I'm a user. I might want my agent to come talk to me, but I'm leaving the cloud or I'm leaving CoreWeave, another cloud. What breaks and what did you guys do to fix it?

>> Definitely. So what breaks today is the cloud agent has limited access to read information about you through the device and also limited functionality to perform actions for you on your device. Right. So they're limited and confined to the applications that they exist in. However, your device today sitting here has a microphone and a camera that understands that we are sitting here and conducting an interview for theCUBE. Right. Your cloud agent doesn't have that access unless you give it that permission. Similarly, your device can take actions for you that a confined cloud agent within an application cannot do. So that's what breaks. And what we really focus on is trying to eliminate that gap. So we build context layers, harnesses, and agents running on the device to actually fully read this context that is critical for personal intelligence and also expose different functions that are missed.

John Furrier

>> How do you decide edge or cloud?

>> Yeah, fantastic question. So there are certain cases where you have to do edge versus cloud. For example, low connectivity, data privacy, ultra-low latency, right? These are instances where you need the edge because a network hop to the cloud doesn't make sense. There are other areas where it's really use case specific, right? And so some cases you could run things in the cloud. But for example, if you're going to ask a cloud agent a simple question such as what's 1+1 equal to, that's going to cost you more money than the economic value that query drives. So you should run at the edge. So really we balance several different heuristics for routing between edge and cloud. We bake that logic into not just how we build our edge models, but how we route between the two.

Dave Vellante

>> So does that make governance and security harder or is it easier because it's more constrained?

>> Fantastic question. Largely it's easier when it comes to data governance and security. So you don't have to contend with where the data goes, access controls, because everything lives on the device. And so that makes a lot of things fundamentally easier. And you can design from first principles your system to decide what functionalities happen on the edge versus the cloud with privacy and governance in mind.

Dave Vellante

>> How do you, how are you working with CoreWeave? Tell us more about your relationship there.

>> Definitely. We as a foundation model lab, even though we train these models that run outside of data centers, we still train on compute clusters in data centers. And our training needs are similar but actually a bit specialized versus how other foundation models train because we are training these ultra-specific alternative architecture models for compute outside the data center. So CoreWeave's been a great partner for us to build the bespoke training stack that we need to build the models.

Dave Vellante

>> And that's your use of CoreWeave, 100% training essentially, or?

>> Largely, yes.

Dave Vellante

>> Yeah.

John Furrier

>> All right, so what's your focus now? You're the COO, you got a great team, you're a lab, you got research, what's commercializing? You got the Snapdragon event, you got the edge, what products are out there? What's the key things you're working on next?

>> Yeah, fantastic question. We're really focused on a series of generations. So where we were when we last spoke is we were building the world's most efficient foundation models. Where we are today is we are building edge agents with those models. So we have a public announcement with Mercedes, we're deploying edge agents in their cars to take action on your behalf. Where are we going from here, right? From here, we want to build these systems and these agents to be able to self-heal and improve and personalize on their own over time. Without that capability, the agents we ship today might be amazing, but AI moves so quickly. Who knows if they're still going to keep up with the pace of user expectation in 6 months, 1 year, right? But why make that a problem that, why can't we bake that, a solution for that problem into the product from day 1? And so we're building observability loops and continuous improvement loops that will improve both the model and the harness over time through natural usage.

John Furrier

>> one of the things we were riffing on with you guys in the past was the edge. Hyperconverged edge, wireless, wireline, putting intelligence at the edge. What is the big bottleneck now for the edge deployments to go full scale? Now we think 2027 will be a big year. Yeah, you got automotive, that's a check. What's telecom sitting there waiting for? They're kind of scared, but once one penguin jumps in the water, they all jump in. We think that's going to happen at MWC this year, we hope, because they're all waking up to the fact that Starlink's around the corner.

>> Yes.

John Furrier

>> Telecom companies sitting on a goldmine of data.

>> Yes.

John Furrier

>> And this is— they're— I don't know what's taking them so long, but what are the blockers that you guys see that you need to either plow through or go around?

>> Yeah, I think it's quality and integration. There are two things, edge models, they're efficient, they can fit, but they— it's well known that when you have a smaller model, you have less capability. And of course, we try to push this envelope, be leading in terms of maximizing quality at the edge. But we don't want anyone to deploy a subpar user experience, right? That's the number one thing for any industry is when is the soonest and how quickly can Liquid accelerate that timeline to bring production quality to the edge? So that's the first thing. And the second thing is integration. Today, the systems that are waiting for their first edge AI are not deploying Edge Foundation models, right? So there isn't an existing Edge Foundation model to swap out. This is often a migration from a system that is not using GenAI today. So as a developer of these foundation models, but also as a company that takes the responsibility to accelerate the adoption of these technologies to the edge, we are also thinking critically about how do we accelerate the integration of these technologies and the migration.

John Furrier

>> All right. So the deal with Qualcomm shows that you can go low footprint, footprint size. A phone's a phone. We know what that looks like. in the telecom world or edge, there's categories of a beefy edge, maybe 30 megawatts, and then central office might be 100 kilowatts. So you're already kind of going from the phone. When you get to the 100 kilowatts to say 500 kilowatts, that becomes a nice edge. What does that look like for Liquid AI? Because if you do it on the phone, I'm imagining that you're going to have a nice purview of a somewhat beefy edge.

>> Yeah, fantastic question. I think as both chips get stronger and energy and wattage goes up, but also as our models get more efficient, what the edge is will push out. And I think that's what you're saying. And we absolutely agree. Yesterday, the edge was a server rack that you put inside a home office. Today it's a phone. Tomorrow it might be a wearable or a network device, right? Space. Exactly. Space as well. So what we are focused on is building a portfolio of models. Even though we build small models, within our portfolio, we have a 100x range of models. Almost 1,000x range between our smallest and largest model size to target the full distribution of edge per industry. And within each of those models, we think about what is the device class that is best suited to adopt this? What are the use cases they care about? How do we proactively tune our models to be great at those use cases?

John Furrier

>> How much functionality do you get with the bigger footprint? 100 kilowatts is a central office. There's literally a $1 billion upgrade path for AT&T to try to put intelligence in there.

>> Right. The more compute you have, the more intelligence you can do. Everything, everyone understands that. I think what we want to do is as you have more compute, so we want to maximize how much production quality you can pack in.

John Furrier

>> So more context.

>> Exactly.

John Furrier

>> More stuff you could use off your small footprint.

>> Exactly.

Dave Vellante

>> Interesting what you said about the spectrum of your models because the challenge with edge from a TAM perspective is huge TAM, but very fragmented.

John Furrier

>> Disaggregated too.

Dave Vellante

>> But disaggregated, but it sounds like you've essentially solved that problem with sort of diversity and optionality of these models with essentially a software business model.

>> That's what we're trying to do. And we don't do it alone. We work very closely with our silicon partners, with our design partners to make sure that everything that we're building at the fragmented edge is very grounded in the needs of that industry, both historically and future-facing.

Dave Vellante

>> Well, what are you using? Different types of silicon, different silicon partners. I think about Sid at d-Matrix and some others, but who are your silicon partners? Obviously, you're using NVIDIA, right?

>> So we train on NVIDIA chips, we train on AMD chips and GPUs. AMD is one of our largest investors, and we work very closely with them across not just data center but also, uh, edge devices such as PCs, etc. We also announced this partnership with Qualcomm focused on the consumer devices all around us. So we really focus on a variety of, uh, compute. And I named a couple large general-purpose silicon providers. We also have partnerships with more specialist vertical silicon providers. Ultimately, what we care about is where does the workload need to run and what can Liquid do to make sure that the technology we ship maximizes quality on that chip and gets input from that edge device to feed back in.

Dave Vellante

>> Exactly.

John Furrier

>> Jeffrey, great to have you on. Say hello to Ramin and the team. Congratulations on the deal.

Dave Vellante

>> Thank you.

John Furrier

>> Close us out. Give a quick plug for what you're looking— people are watching. How do they engage with you guys? What are you looking for? Hiring, funding, put a plug in for the company.

>> Definitely. Liquid AI, we are— we have big ambitions. If you are interested in working on foundation models outside of data centers, check out our careers page. We also just released a very interesting model yesterday. It's our first vision model and is actually the first model to beat Gemma at its own game. It's available online. So if anyone wants to try that out, give it a whirl and reach out if you have any questions.

Dave Vellante

>> How many are you at today?

>> Today we're about 120 people.

Dave Vellante

>> Oh,

John Furrier

>> all right.Sizeable company. And CUBE alumni now, and of course NYSE Wired community members. Thanks for being part of the

>> team.Of course. Thank you for having us.

John Furrier

>> All right. We're doing our part here. We're live streaming from the show floor in San Francisco. The AI generation is about intelligence distribution, creating intelligence. AI factories create intelligence. Now they're centralized. They'll go to the edge. They'll be in devices. That is distributed computing that we're doing our part here on theCUBE. Thanks for watching.