CoreWeave Fully Connected 2026 | Renee Huang, Cline

Clips
More from CoreWeave Fully Connected 2026

Renee Huang

CPO

Cline

play_circle_outline Cline: Modular Open-Source Harness for Agentic Coding on Open-Weight Models with Evaluation, Failure-Correction, and Self-Hosted Enterprise Deployment

play_circle_outline Open Models, Local Fine-Tuning, and Routing: Balancing Fast Small Models, Frontier Trade-offs, and Plugin Integration

play_circle_outline Agents, semantic layers, and knowledge graphs provide contextual data for models

play_circle_outline 20x Inference Token Surge Since May and CoreWeave Partnership for Serverless Inference Pricing

Info
Transcript

Renee Huang, Cline

Renee Huang

CPO Cline

In this interview from Fully Connected 2026, Renee Huang, chief executive officer of Cline Bot Inc., joins theCUBE's John Furrier and Dave Vellante to discuss how open-source coding harnesses are expanding access to AI development. Huang explains how Cline's fully open-source codebase gives individuals and enterprises the flexibility to self-host or customize their deployment, tailoring the harness to open-weights models rather than locking users into a single frontier model. She details the rigorous evaluation system built over two years to ensure the harnes... Read more

explore Keep Exploring

What does your company do, and how do you differentiate your harness from other solutions—especially regarding being open-source and supporting open‑weight models? add

How should model routing be handled between smaller, lower‑cost/open‑weight models and larger frontier models—is that a key differentiator—and what developer use cases are emerging for open‑weight models and inference (e.g., local/private deployment, fine‑tuning, and harness/plugins)? add

How are traditional data practitioners and newer AI-focused developers converging, and what are the trade-offs between using a semantic layer versus a knowledge‑graph architecture to provide context for AI? add

What is the relationship between the company and CoreWeave? add

bolt Powered by CUBE AI

Renee Huang, Cline

search

(INTRO)

John Furrier

>> Welcome back to theCUBE's live coverage here in San Francisco for CoreWeave's first annual conference. It's called Fully Connected. I'm John Furrier, your host with Dave Vellante. Getting all the action on the AI infrastructure build-out as the new layer of infrastructure emerges, more production systems of intelligence are emerging that will fuel the AI-native wave that's continuing to grow. Our next guest is in the middle of it. Renee Huang is here, CEO of Cline. Welcome to theCUBE, appreciate you coming on.

>> Thank you so much, John, Dave.

John Furrier

>> You have a talk coming up too here at the event. This has been really kind of an interesting moment with CoreWeave because we saw this when we went to the first AWS re:Invent in 2013. Early players are in.

>> Yeah.

John Furrier

>> Real actors and participants in the innovation. Ecosystem's forming. Use cases are emerging.

>> 100%.

John Furrier

>> Line of sight on profit and revenue. The tech is evolving super fast.

>> Crazy fast.

John Furrier

>> Explain what you're doing because coding has opened the door to the enterprise. Chat and retrieval, it's kind of nice, nice chatbots, nice to have. But the coding really opens up the value proposition to agents, physical AI right behind it.

>> Yeah, totally.

John Furrier

>> Seeing a lot of action. What are you guys up to? Explain what you guys do.

>> Yeah, so I lead a company called Cline and we are an open source coding harness and actually evolving into a more general harness and we tailor very specifically to open weights models. So in the early days, like we are among the first to introduce that agentic coding experience to people and that so far has really been unlocking the whole productivity gain starting from developers. But you are really seeing that penetrating into people who don't really code day in, day out. and also people just doing knowledge work, but really giving the toolkit for them in a very open format on what they can achieve and what they can do on a day-to-day basis.

Dave Vellante

>> So there's a lot of harnesses out there, obviously. So talk a little bit more about how you differentiate. I get the open and open weight, but maybe you could double-click on that.

>> Yeah, totally. So, I think we start with our whole codebase is open source, so we give that transparency to people, and that gives a lot of deployment choice for both individual user but also enterprise. That if they don't want to use some managed solution, they could just take our solution and host it themselves. So that openness and transparency is kind of in our DNA from day one, very similar to the open-weights model story. Now, for a harness layer, to your point, yes, there's a lot of different harness and the harness constantly evolving as the model capability really gets better. And for us, it's a whole kind of muscle build out throughout these 2 years. To really have a very rigid type of evaluation system from what we see from the real-world traffic, what type of use case people are doing that's getting more and more complicated day in, day out. And how do we design a very modular architecture behind the scenes to give the model the right tool, the right correction loop? How do they detect failure and how do they fix themselves? What is the right type of different instruction for specific model class? Regardless if it's larger, more frontier type or smaller that people just want to use it for faster and cheaper type of inference. So we really try to capture the whole spectrum with a very rigorous style of evaluation.

Dave Vellante

>> What are some of the things that your customers are doing? Maybe you could give us an example of sort of your ideal customer.

>> Yeah. So we serve across the user base of individual developers because we want to give the greatest access to everyone, that it doesn't have to be a large enterprise. So a lot of individual users just use us for very easy access to all different open-source models because there's new ones coming out every week, 2 great ones coming out every week, and they just need an easy entry to all those models with a good harness that's wrapped around it. And we really try to be that platform. And then from an enterprise perspective, we see people just building upon us all the time. And the type of architecture we offer to people is very modularizable. People can just plug in things to it so they can have customization upon it. And that flexibility that we really see benefits across the spectrum for everyone.

Dave Vellante

>> So I know developers and AI engineers, they'll say, "Oh, I love Claude, but I need a better choice of model. I need" lower-cost models, and that's where open weights comes in. So, how do you deal with model routing and things of that nature? Is that a big differentiator for you? Because more companies are doing that now, but maybe you could explain.

>> Yes. So we haven't yet been deciding the model routing for people yet. What we've been doing really, really well is to have our harness that's performing good, top tier, honestly, both on the smaller Flash model and on the larger, more frontier model, so that a user intentionally coming in with that trade-off in mind— do I want a very complicated model for the most difficult task today versus do I just want something that's more cheap, economic, and faster— so they can intentionally make that decision themselves. We are thinking about routing very heavily internally. And it's just really like we want to do it the right way.

Dave Vellante

>> But out of the box, you can actually satisfy that wide spectrum.

>> Yes.

Dave Vellante

>> As you suggested. So that's the appeal of Claude.

>> Totally. Totally. That flexibility and choice.

John Furrier

>> You're giving a talk here. What's the topic of your talk?

>> We are really trying to promote the whole idea of open source inference with open weights models. And really for us, the mission is really to give access, best access to everyone. Right? We don't think AI should only be given to the people who can afford it, paying at like very premium subscription, like being at a large enterprise who have unlimited budget. We really think AI should belong to any type of day-to-day work. And we've seen this trend of open-weight model actually just totally taking off consumption in our platform. And that echoes to the market demand of, hey, people really want that choice and people really want that easy access.

John Furrier

>> We love— we think there's going to be a real boom for open weight models and inference. Obviously, if they're done right.

>> Yeah.

John Furrier

>> Because there's a real need for specialized intelligence and agentic workloads are going to be intelligent. You mentioned some of those things.

>> Yeah.

John Furrier

>> What are some use cases that you're seeing now that are emerging from the developers? Because they're the canary in the coal mine.

Dave Vellante

>> Yeah.

John Furrier

>> They're going to tell us what they're doing. What are they doing? Is it small models? Because you can take an OpenWeight model, you don't have to get all the weights. Great, I can craft my own, I can put my ingenuity.

>> Yeah.

John Furrier

>> And develop. What are some of the key use cases?

>> It comes on both the model layer and inference layer. On the model layer, there are people who are extremely privacy-focused who just want to do local, even local model development. Then now like there are good enough small models, they can just install in their laptop and running on local GPU and even get inference. So there's a lot of customization and some people even do their own fine-tuning and post-training on it. On the harness layer, there's a lot of different plugins we call it, which are additional tools that you want to customize upon a harness. And that people do that to obviously the obvious idea is to connect to more different toolkit. Of their own proprietary intelligence. But also, for example, if I'm trying to specialize in a backend task, or if I'm trying to specialize in a more physical AI 3D style task, I'm going to give it some special type of tooling that it can achieve it better.

John Furrier

>> Renee, we like to study all the hype cycle and try to get that signal out of the noise. And one of the things that's come up is, for lack of a better, phrase, it's pejorative, citizen developer. But there are small things that people do. And there's also a perception of I need to have all these machine learning engineers to do all this work to build a model. You're kind of opening the door for anyone to just do development. But what's all the hype now is another term called asset light.

>> Asset light.

John Furrier

>> Asset light, meaning I'm not going to pay all the CapEx.

>> Yeah.

John Furrier

>> I'm going to put it on a DGX box or on a localhost, build on my laptop. Laptops are getting better. Dell's next laptop's gonna be pretty massive. NVIDIA's got these amazing machines.

>> Yeah.

John Furrier

>> For like $5,000, you can run a lot of stuff on. So you're seeing entrepreneurship and creativity.

>> Yeah.

John Furrier

>> Booming at this phase. This is not wrapping. So that was to use the proprietary frontier models, but now I can grab an open source model for free.

>> Yeah.

John Furrier

>> Grab the open weights and I could target a problem that I want to solve. That's what we're seeing. That's kind of the signal. What are you seeing there? Because that breaks the conventional wisdom of, okay, let's do a deal with a Frontier or let's buy some GPUs. So here with CoreWeave, they're large scale. So they're already kind of— Yeah, the mark on— we got all the GPUs, everything you need. It's a full ecosystem. Chen Goldberg's like, okay, we got all that. They got all the orchestration, they got a bunch of Kubernetes in there and all the stacks, perfect. So if I'm a developer, I'm looking at CoreWeave like, hmm, I want to just win first my idea.

>> Yeah.

John Furrier

>> Then go to CoreWeave. So I can see how CoreWeave might be interested in nurturing that market. Am I getting that right?

>> I think it's always a trajectory, right? There's always going to be a scale-up phase once the idea is right. And exactly to your point, I think the benefit of all those open tooling, including open weights models, are really lowering the barrier, the entry barrier of, hey, I just want to test out something. But then always as you grow up, you have to mature and then you have to get to a larger platform. All those are still very important problem to solve, but you are capturing a much larger radius right now.

John Furrier

>> Okay, so here's a question for you. So I'm a developer, I'm coding away. I say, hey, theCUBE has a big corpus of vector data. I'm going to go code something. I don't have a lot of budget and I'd lock in, oh, for a week I code something. Oh my God, the model's changed. There's a lot of, you mentioned that earlier, there's a lot of different models changing.

>> Yeah.

John Furrier

>> What happens to me? I'm now, I'm like, am I paranoid? Did I use the wrong engine? What am I building, what did I, so this is a little bit of cognitive dissonance, but I'm just bringing it up because I've seen this before where it's like, wait a minute, I did everything on this model. Kimi was great, but what about this over here? So how do you handle that? What do you see developers do from a practice standpoint?

>> That's a very interesting question. I think people going into a problem, there are two types, right? One, you're just trying to do open-ended exploration, and two, you have some type of goal in mind that you're trying to achieve. So I think for the developer that we are working with and really for a lot of the serious developers out there, they are typically going in with a goal that I want to build this type of thing. I have this testing guardrail that I want to achieve and I'm making sure this model that the task it accomplishes, it can kind of meet all those criteria that are listed out. And that are how we are using how all the, I think Model Lab and our Harness development that we use a lot of evaluation, which is essentially making sure that the task people want can be achieved as they intended.

John Furrier

>> I was talking to a large enterprise, Dave, and they said, hey, we have Amazon, we've been in the cloud for over a decade. We have a relationship with CoreWeave. We have some GPUs in the sandbox and we're doing some cool stuff. But some of the use cases are long-tail use cases.

>> Exactly.

John Furrier

>> And they go, well, I really don't want to throw a lot of money at that. And so they say open weights have been very, very important to them.

>> Yeah, yeah.

John Furrier

>> Why are people excited by open weights? Besides the fact that it democratizes, it lowers the access point to get in, is there a real value on that power law? Because you could achieve the task and then move it up into production or other things. Why is open weights so important right now in your mind?

>> I think your point is exactly the longer tail ones you want more control and customization. I think really the customization again falls back to both the model layer itself and also to the harness and tooling layer. You don't really want to pay a huge standardized SaaS budget if it's like a long-tail problem. You don't really want to hire tons of engineers if it's a long-tail problem. But now, because of how flexible things are, you can just build it.

Dave Vellante

>> I'm checking out your GitHub. Cline, the Cline GitHub, it's pretty impressive. You've got 352 contributors, different from committers, but still that's good. And then, but almost 70,000 stars.

>> Yeah.

Dave Vellante

>> So, you got some forks going on and some branches, but that's pretty impressive. it's not, you're not a household name yet, but maybe that's your ambition. Maybe you feel like you are, but very popular. those are pretty good ratings.

>> Thank you.

John Furrier

>> Thank you.

>> Yeah, we run a lot of tokens through us. We were one of the very early open source harnesses. I think we were probably the fastest trending GitHub repo last year. Yeah.

Dave Vellante

>> Yeah.

John Furrier

>> Right, right.

Dave Vellante

>> Very cool. So what is your ambition for the company? When you founded it, what were you thinking and how has that ambition changed and what's your vision?

>> Well, yeah, we are just so excited. When we first built it, it was really like there were only a couple closed models that were good. And now it's almost like we are cooking with way more ingredients that's in the kitchen for you to utilize. So yeah, our mission hasn't changed from the beginning, which is to really try to give the easiest access to as many people as possible. That's why we use a developer-first approach. Now with the agents getting better, with the model getting better, we are seeing this penetration trend into more of the general knowledge work, into people who are less sophisticated in coding tooling. And we introduce different surfaces that can be more helpful to them. Previously, it's more of an IDE or a terminal line, command line that you have to feel comfortable about some of those tooling. But now there's more native application layer. There's more like mobile app that you can just really access work from anywhere.

Dave Vellante

>> And what's your background? Were you the first committer?

>> No, no, no. So I've been with Cline for about a year. I joined from outside. But yeah, it has been a real blast. A blast, a whole blast to see the journey so far.

John Furrier

>> What's it like for these two generations of data people, which we are one of them, we're data hoarders, we like to have data, we cover the data platforms. In the old IT days, it was the data lakes was an innovation. Yeah, that seems like not a big innovation compared to what's going on in AI. The young guns coming up are merging in with the old school infrastructure data folks. What's it been like to see those worlds collide, intersect.

>> Interesting. Yeah, no, I came from data world as well. I was at Snowflake for a while. So I think it's really like that data plus AI. Data is a context layer basically, right? And then you got this superpower that's upon it that can achieve more use cases. And I think that's for all the individual users, but also for large enterprises as well, putting your AI with the right context. And being able to meet different spectrum of use cases. That's eventually—

John Furrier

>> a lot of developers love the semantic layer because you can define the business, but the knowledge graph— graphs have become the relationship maps. Yeah, which play well in models. Yeah, we see developers kind of riffing on that, building these platforms. What's your take on that kind of architecture?

>> Interesting. Right now I agree with you, there's no one direction that you can tell is better. I can see the benefits of both. And I think eventually it probably depends on your task and the complexity of the data relationship that you are trying to fit into the context layer.

Dave Vellante

>> What's your relationship with CoreWeave?

>> So we use a lot of their serverless inference. So like I mentioned, as OpenAI is getting really good, actually our token consumption on our platform has 20x'd over the last 4 months.

Dave Vellante

>> 20x?

>> 20x. It's crazy. On all the OpenAI models.

Dave Vellante

>> Predominantly inference or?

>> Yes, token consumption through inference.

Dave Vellante

>> Yeah.

>> User inference.

Dave Vellante

>> It sounds like that curve is steepening or is that?

>> It's definitely like, there's definitely an inflection point this May. I remember very clearly because that's when There are a couple of good open weights models out there and people are getting really excited.

Dave Vellante

>> How do you purchase those tokens? Are you making long-term commitments? Are you doing on demand?

>> We are still trying to figure it out. And that's part of our relationship with CoreWeave. So we leverage a lot of their inference under the hood. And I know they are coming out with all different types of pricing, provisioned throughput, flat reserve. So We are figuring out.

John Furrier

>> A lot of innovation. Obviously, Weights & Biases was acquired.

>> Yeah, yeah.

John Furrier

>> Nice to have Lukas on the team there.

>> Yeah, totally.

John Furrier

>> So very cool. And you got your talk coming up. Close us out by sharing what you're specifically going to talk about in the talk. Is it going to be technical? Is it going to be more directional? What's the topic? Is there a link on your LinkedIn?

>> Yeah. Yeah, yeah, yeah. So it's going to be a mix of more directional and technical. Obviously, share the trend again, what we see from our consumption pattern. How people really love the open weights and what type of task distribution people are trying out. But then also trying to bring more into the topic of why does a harness matter and what are we trying to make easy for our user so that they don't have to worry about building everything.

John Furrier

>> What a great innovation. Renee, thank you so much. Have a great talk. Sorry.

Dave Vellante

>> Good luck.

John Furrier

>> Thank you.

>> Thank you.

John Furrier

>> I'm John Furrier with Dave Vellante. Again, the AI wave is here. Again, the harnesses are going to be a big part of the agentic. Not just MCP servers, but the data is the context. You got to feed the systems that are producing the intelligence. Of course, the developers are going to have a feeding frenzy over this. We'll be watching it closely. Stay tuned for more after this short break.