RAISE Summit '26 | Lin Qiao, Fireworks AI

Clips
More from RAISE Summit '26

Lin Qiao

CEO & Co-Founder

Fireworks AI

play_circle_outline Scaling AI in the Enterprise: Infrastructure, Production Readiness, and Proprietary Data as Strategic IP

play_circle_outline Specialized intelligence vs general intelligence (AGI) explained

play_circle_outline Harness/orchestration for managing agents, workflows, and model calls

play_circle_outline Open vs Closed Models: Control, Customization, and 5–10x Cost Savings from Tuning Open Models

play_circle_outline Fireworks: Speed, Scale, and Tooling for Building Specialized AI Without Scarce Expertise

Info

Lin Qiao, Fireworks AI

Lin Qiao

CEO & Co-Founder Fireworks AI

search

>> Palo Alto Studio Connection, Silicon Valley and Wall Street.

>> I'm John Furrier, host of theCUBE, here with Gemma Allen, my co-host.

>> Welcome to the RAISE Summit coverage, brought to you by Solidigm, with support from Tensordyne, d-Matrix, and Argentum.

>> Welcome back to theCUBE's livestream here in Paris, France. I'm John Furrier, your host of theCUBE. We're here for the RAISE Summit, day 2 of live coverage. The conversation continues. The AI infrastructure buildout is enabling a lot more adoption, real-world use cases. And as the enterprises and application developers and AI-native startups are developing great apps, the need for models continues to thunder away and change. A lot of debates, a lot of conversations around intelligence, general, specialized.

>> We got a great guest here.

>> Lin is back on theCUBE, CEO of Fireworks and co-founder. Great to see you. Thanks for coming on theCUBE.

>> Appreciate it.

>> Thanks for having me.

>> So first of all, congratulations on all the momentum with Fireworks AI. You guys are doing a great job. You were on theCUBE twice. Last time we talked, we were talking about adoption, the enterprise stack, and we were just starting to see the momentum.

>> Yes.

>> Now you're starting to see people leaning in because coding has opened the door. Yeah, we had search, we had good marketing copy with the model OpenAI and Claude. But now when you start getting code generation, they see value. Now we got the agents coming in. Inference is eating the world. Now you're starting to see people really think through, how do I use AI and inject intelligence into my organization? This is something that you guys have been very, very successful at. Give us your thoughts on this because now you're starting to see the adoption, not just kicking the tires, seeing real workloads. What are some of the issues that you're seeing? What's the adoption look like? Share your thoughts.

>> Yeah, so it's interesting over the course of years, especially starting from last year, the model quality has significantly leaped forward.

>> Both open model and closed model combined have passed the threshold of the quality. It can solve a wide range of tasks as is.

>> So that's why we start to see adoption across the industry.

>> Many, many companies, whether they're startup digital natives or enterprises, are all investing in AI. Very aggressively, and actually it's not a prototype anymore. It is actually going to production scale.

>> At the same time, because the quality passed the threshold, it makes the open model extra appealing because of multiple reasons.

>> One is the model quality is so good, then after you tune it, it can be extra good. Extra good meaning it can beat the Frontier Labs off-the-shelf models to solve your domain-specific problems or solving your application-specific problems. And the fundamental reason is there are multiple drivers why we have so many customers want to own their own intelligence and have it under control is existential because you think about intelligence as a core IP of running your business. You will never give away that intelligence to another company or to your competitors. Right. Does that make sense?

>> Yeah.

>> So But if you build on top of off-the-shelf black box API, then you're pretty much giving away everything. That is critical, the core part of our product, because application before was a moat. Developing application, having that intuition, implementing it, production scale, it is hard. You need to have a strong team, sizable critical mass of product engineers, PMs.

>> Now one person, multiple weeks, they can do it. They do not need to know how to write a single line of code. They just have brilliant ideas and they can test it quickly.

>> That significantly collapsed the resource and talent requirement of a state-of-the-art new application launch. And the application by itself is thin in terms of the moat.

>> So what is the real moat? The real moat is the data collected from your application for deeper understanding of user preference, user intent, why users engage, why they do not engage, and use that data to continuously improve your product.

>> And that is intelligence. And today, majority of data globally are those data, are those private data locked inside application, locked inside enterprise.

>> It's very interesting.

>> We firmly believe, I think all people, many people agree with me, is intelligence is derivative of data.

>> And today, majority of data is unseen.

>> Majority of data are those private data never being activated to create any intelligence.

>> And that's where we believe the new frontier is the specialized intelligence belonging to individual application, individual enterprises. Those are next frontier we are pushing on.

>> So Fireworks is a specialized intelligence platform We enable every single company for every single use case to have their specialized model and specialized intelligence. Generalized intelligence, usually you hear the AGI companies, those are parallel strategies and we believe both will coexist, but the future will not be dominated by a few models in the AGI category. We believe future will be millions of specialized intelligence models.

>> Yeah, there's

>> different—explain general intelligence and specialized intelligence. People are not—

>> may not understand the difference.

>> Explain that. General intelligence versus specialized intelligence.

>> Yeah.

>> So you can think about general intelligence as the process of learning from common tasks that we humans perform.

>> So for example, analogy is when we were born, All of us have IQ that doesn't change, right?

>> So we human beings species, our IQ is way higher than the rest of the planet species. So that is when a generalized intelligence model was born, it carries the fundamental IQ that solves a lot of tasks, common tasks, because it learns from public internet where, by volume it knows how to solve this problem, that problem, and so on. So that's a general intelligence model, and you just use it off the shelf as is. But specialized intelligence are domain-specific problems or product-specific problems that only the product designer, they will create that. There's no way a general intelligence model will ever be able to learn from it because it has tons of in-product context to optimize it. For example, we— my background is we worked many, many years at Meta. One type of AI problem is recommendation, right?

>> But recommendation is personal.

>> Every single person will have their own preferences based on their interests, based on their background.

>> Based on their region. And there's no uniform way to predict what you like, what I like. And we all want to see different things, different video recommended, different product recommended, different newsfeed recommended. So this is just one example.

>> Obviously, there's a ton of other examples. So this type of personalized recommendation is just one form of specialized intelligence. But there are many other forms of specialized—

>> people can rationalize, okay, use OpenAI, ChatGPT, Claude. Hey, I want to do a task. They give it a prompt, they get an answer. Maybe it reasons, multi-step reasoning.

>> Okay, great.

>> But the enterprise, OpenAI doesn't crawl that. So again, it's locked.

>> Yes.

>> So the unlock of the private data becomes a big thing that's specialized.

>> Right.

>> And workflows are also not applications. Applications are like, hey, I built an app. And you ship it, put it in the App Store. Workflows are how companies run their business. And that has data behind it. And it's behind the firewall and it's known to them. That is what I think specialized is.

>> That's specialized. And the other way, a form of very popular specialized intelligence is the orchestration of your workflow and the orchestration of how you call into different AI models.. And that orchestration today is called harness.

>> It's specialized, it's customized, bespoke. And that harness usually is also powered by a model behind, and that logic is specialized towards this particular application. And who designed this? There's no common knowledge of how to orchestrate this. So those are all examples where you, the product design or a specific implementation design, need to co-train with the model so model really learn what you want to optimize for. So there are gazillion number of those examples where just a general purpose is not good enough and you need to have specialized. General purpose doesn't give you control.

>> Yeah.

>> Specialized gives you full control based on open weights. That's where the significance of open-weight model is here.

>> And you have this proprietary intelligence powering your product, make your product better. Your product drives more data, make the model better, better model, better product.

>> And then you start to play this.

>> Lin, you mentioned harness. I love that term. It comes up a lot. Harness is a mechanism for managing agents and tasks, let agents do their thing. When you think about open-source software, we love open source. We use all that. It's open-source software. You take it and you modify it.

>> And that works.

>> That's been generations of success have been built on open source. So what you're saying is open models, you take it and you modify it. This is why OpenWeights is so important because people don't understand that if you can use the OpenWeights, you can tailor it to your needs. Like open source software. I can get open source software.

>> I can add to it.

>> Now, the difference is you can keep the weights inside the company. This is a huge enterprise value proposition. So how are people using open models? What's your experience? You got a lot of customers.

>> You guys are the leader.

>> How are people setting up the harnesses? What are they doing with the open model? Because it makes total sense. I got data, I got engineers, I take the model. How do— how are people taking the open models and tweaking them?

>> Yeah. So first of all, I think end of last year we see a huge uptick of leveraging open models, leveraging open models as is, because again, the quality has passed a threshold.

>> And second is significant interest in customized open model to turn that into their own model, not one per company, one per task across a wide variety of tasks. So in a way, actually, the whole entire, I think, industry starts to learn quickly how to manage open models.

>> Exactly as you mentioned, the benefits are full control.

>> You can change all the weights, but it has a really solid baseline for you to build a delta on top of.

>> It's so much easier to tune.

>> And second is it's much cheaper.

>> So the open models, the cost can— you can bring down the cost.

>> Everyone's doing token maxing, right?

>> And then they look at the— they're carefully looking at the bill.

>> It's like, what is ROI here? Can you justify it? No one can. So you can bring down the cost without losing quality by 5 to 10 times. Massive, significant cost reduction.

>> And these two values combined are creating a huge explosive adoption of open models.

>> So the way to do that is first you need to define a judgment of whether your tuned model, how good they are.

>> So this is called eval. People start to create evals and evals are very familiar concept in general software development because we all need to write unit test, integration test, right?

>> So there are a lot of kinds of testing, and it's just a form of testing quality. And it is not different.

>> It's like software development.

>> Yeah.

>> And obviously, these tests need to continue to be more comprehensive.

>> But the key point is, once you have these evals, and then everyone has data. And then we can help you hill climb. Hill climb means you just optimize for those evals. And it's the more we have so many cases again, we can get there. And many leaders already start to pioneer this even from 2 years ago. Now, Cursor is one of our flagship customers and they just launched a new model yesterday, actually. Yes, Grok 4.5.

>> So they have been leveraging their data to tune the model.

>> Specially designed for the coding use cases.

>> Really, really good. And we have been partnering with Harvey, for example.

>> They establish a legal bench that is their eval, and we help them climb towards that benchmark.

>> There are many, many other examples. I think this year we also see— last year was the coding year.

>> This year we see a lot of the co-worker applications, there are general purpose co-workers, there are specialized co-workers across healthcare, insurance, legal, recruiting, marketing, sales, customer support across the board.

>> Very interesting.

>> And they all have their internal evals because if you want to really be a company, eval is the first step. Without eval, you cannot really do hill climbing.

>> Hill climbing.

>> Yes. What's the next progression?

>> And hill climbing is actually we provide a variety of tools.

>> With all different kinds of algorithms. Obviously, the most well-known one is actually reinforcement learning.

>> The interesting thing is for people to do reinforcement learning, our philosophy is we anchor on the best model. So we do tuning on the best, highest quality model to hill climb there. And then once,

>> because that's—Explain what hill climbing is first. I love

>> this.Yes. So hill climbing is literally like you climbing a mountain, right? So basically you say,

>> hey, I—Get in shape.

>> I want to climb that

>> mountain.That's the peak I want to get to.

>> The peak is our eval. And here, it's a training process.

>> And the training process basically says, this is my objective. And the training is going backwards, forward, backwards, forward in order to optimize for that objective and then get there.

>> And sometimes it cannot get there, it collapses.

>> Then, this

>> fails.Fall down, you get back up.

>> Yeah.You get back up and you change a bunch of things and try again.

>> So that's the analogy we use a lot. So after you have done that, and then you will get a good model, and then we use—

>> we immediately optimize for inference. So our platform is training inference co-design, co-optimized.

>> So because at the end, training is not the endpoint of this process, it really needs to deploy to inference, let your product use it. And then you do—

>> a lot of people do A/B testing because they compare A and B with or without this new model and see whether people like the product change or not.

>> Right.

>> So this is part of experimentation. Once they see strong product feedback, then they scale it.

>> So we are just one platform for you to do everything together.

>> And our job is to make it super easy that everyone can have specialized intelligence model.

>> And our vision is the future.

>> There will be millions of them.

>> And people are like, this is way too big.

>> But we really believe that. We really believe every single app company, every single enterprise, they deserve to have their own specialized model.

>> It can be scary for folks to see the big picture. I want to get to what your platform does because I think this is super important. You have two schools of thought. I want to go fast or I want to build my own.

>> Yes.

>> And you guys work with Harvey, you mentioned Cursor, you have a lot of the marquee AI companies. What's the reason why they use Fireworks? Is it because they want to go faster? You guys are providing the tooling, right? the skeptics say, well, I don't need Fireworks, I just do it myself. Explain the value, because I think there's a lot of tooling that you guys are providing.

>> There's evidence.

>> Yes.

>> Talk about that value, but why Fireworks?

>> Yes.

>> What's in it for me? What's the benefit? Right.

>> So I had an interesting conversation with Jensen after his GTC keynote. He mentioned one thing. There's no such thing as a general company. There's no such thing as a general company.

>> Every company exists for a reason because they add exceptional value in the ecosystem and they specialize really well. On a certain dimension.

>> So this is a product design principle.

>> Same for intelligence.

>> In our view, product and intelligence today, they are not decoupled.

>> They are one thing. So every company needs specialized intelligence. So that's our view.

>> And that's where the industry is going towards because, as I mentioned, it's existential

>> It's faster with you guysIf I want to do my specialized intelligence, I'd go with Fireworks. I'm on board. You're optimizing for that goal.

>> Right.

>> So I'm behind from day one. The catch-up, even then, I'm not optimized.

>> Right.

>> So we want to— we feel the product team or enterprise, they own their data. They own the brain. And we are the hands. For them to implement what the brain wants. So that's how we think about how we help them accelerate this journey. And in order for the brain to operate, to build this intelligence, they need to have deep AI knowledge.

>> They also need to have deep AI system knowledge to manage the whole entire thing.

>> And there's a lot to build. And the resource pool, a talent pool, is very scarce.

>> And we're like, hey, we can build this layer.

>> You build on us.

>> You focus on your biggest differentiation and we help you there.

>> So that's how you think about it.

>> You help them scale too. We have scale as a huge piece.

>> Absolutely.

>> Yes.

>> What's your biggest focus this year as you guys continue to have the momentum? By the way, I love the product.

>> What do you guys do?

>> You get people up and running and you continually add value.

>> To your point, what's your focus this year?

>> So our focus this year is to make our product way easier to use. Hey, we're living in the age of agents. We are in the process of automating a lot of things so our customer can use our product and run really fast. The second is to reach out to many people who are in need of—

>> we see a lot of demand, so we just need to grow our company and scale it so a lot of people can access our product. So those are the key focus, but we really believe we are on the right track of deploying this specialized intelligence to everyone. But as we talk about that, there's a huge debate across the industry about, hey, should there be one company owning all the intelligence or it should be everywhere democratized?

>> I think AI is such an important technology, piece of technology.

>> It shouldn't be owned by only one company.

>> It just feels— doesn't feel right.

>> You're enabling people to turn on their corporate brains.

>> Yes.

>> Basically.

>> hey, and you're not threatening anyone because you're open.

>> Yes.

>> And you're scaling people faster.

>> Absolutely.

>> Yes.

>> when ChatGPT came out, Dave Vellante and I published a very controversial chart called the Power Law of AI Models. It was basically a curve. It had a torso and a neck. And we speculated that we think there's going to be specialty models.

>> And guess what?

>> You're now pioneering the new frontier. And if you look at the data in the enterprise, only about 1% of their data is unlocked. That means that every company is sitting on a treasure trove of data that could be activated and turned on and connected into a global brain.

>> That's what you're doing.

>> That's the beauty of the time. the opportunity is so big here.

>> So that's why we are super excited to help everyone. Get there.

>> Well, Lin, great to see you in person here in theCUBE in Paris.

>> We'll see you in the Bay Area. Thanks for coming on theCUBE, really appreciate it. Congratulations on all the success and keep going.

>> Thanks for having me.

>> All right, the specialized model, specialized intelligence, this is going to unlock the enterprise. Coding was step one. People can see the value of coding, code generation that opens up the value is undeniable when they start seeing the actual data of the enterprise being codified and enabled for AI applications, the brain gets turned on, AI works, the benefits are coming. And again, as the AI infrastructure continues to build, more and more of the enterprise stack will see this capability. We're here at the RAISE Summit doing our part to bring the action to you. Thanks for watching.

Lin Qiao, Fireworks AI

search

>> Palo Alto Studio Connection, Silicon Valley and Wall Street.

>> I'm John Furrier, host of theCUBE, here with Gemma Allen, my co-host.

>> Welcome to the RAISE Summit coverage, brought to you by Solidigm, with support from Tensordyne, d-Matrix, and Argentum.

>> Welcome back to theCUBE's livestream here in Paris, France. I'm John Furrier, your host of theCUBE. We're here for the RAISE Summit, day 2 of live coverage. The conversation continues. The AI infrastructure buildout is enabling a lot more adoption, real-world use cases. And as the enterprises and application developers and AI-native startups are developing great apps, the need for models continues to thunder away and change. A lot of debates, a lot of conversations around intelligence, general, specialized.

>> We got a great guest here.

>> Lin is back on theCUBE, CEO of Fireworks and co-founder. Great to see you. Thanks for coming on theCUBE.

>> Appreciate it.

>> Thanks for having me.

>> So first of all, congratulations on all the momentum with Fireworks AI. You guys are doing a great job. You were on theCUBE twice. Last time we talked, we were talking about adoption, the enterprise stack, and we were just starting to see the momentum.

>> Yes.

>> Now you're starting to see people leaning in because coding has opened the door. Yeah, we had search, we had good marketing copy with the model OpenAI and Claude. But now when you start getting code generation, they see value. Now we got the agents coming in. Inference is eating the world. Now you're starting to see people really think through, how do I use AI and inject intelligence into my organization? This is something that you guys have been very, very successful at. Give us your thoughts on this because now you're starting to see the adoption, not just kicking the tires, seeing real workloads. What are some of the issues that you're seeing? What's the adoption look like? Share your thoughts.

>> Yeah, so it's interesting over the course of years, especially starting from last year, the model quality has significantly leaped forward.

>> Both open model and closed model combined have passed the threshold of the quality. It can solve a wide range of tasks as is.

>> So that's why we start to see adoption across the industry.

>> Many, many companies, whether they're startup digital natives or enterprises, are all investing in AI. Very aggressively, and actually it's not a prototype anymore. It is actually going to production scale.

>> At the same time, because the quality passed the threshold, it makes the open model extra appealing because of multiple reasons.

>> One is the model quality is so good, then after you tune it, it can be extra good. Extra good meaning it can beat the Frontier Labs off-the-shelf models to solve your domain-specific problems or solving your application-specific problems. And the fundamental reason is there are multiple drivers why we have so many customers want to own their own intelligence and have it under control is existential because you think about intelligence as a core IP of running your business. You will never give away that intelligence to another company or to your competitors. Right. Does that make sense?

>> Yeah.

>> So But if you build on top of off-the-shelf black box API, then you're pretty much giving away everything. That is critical, the core part of our product, because application before was a moat. Developing application, having that intuition, implementing it, production scale, it is hard. You need to have a strong team, sizable critical mass of product engineers, PMs.

>> Now one person, multiple weeks, they can do it. They do not need to know how to write a single line of code. They just have brilliant ideas and they can test it quickly.

>> That significantly collapsed the resource and talent requirement of a state-of-the-art new application launch. And the application by itself is thin in terms of the moat.

>> So what is the real moat? The real moat is the data collected from your application for deeper understanding of user preference, user intent, why users engage, why they do not engage, and use that data to continuously improve your product.

>> And that is intelligence. And today, majority of data globally are those data, are those private data locked inside application, locked inside enterprise.

>> It's very interesting.

>> We firmly believe, I think all people, many people agree with me, is intelligence is derivative of data.

>> And today, majority of data is unseen.

>> Majority of data are those private data never being activated to create any intelligence.

>> And that's where we believe the new frontier is the specialized intelligence belonging to individual application, individual enterprises. Those are next frontier we are pushing on.

>> So Fireworks is a specialized intelligence platform We enable every single company for every single use case to have their specialized model and specialized intelligence. Generalized intelligence, usually you hear the AGI companies, those are parallel strategies and we believe both will coexist, but the future will not be dominated by a few models in the AGI category. We believe future will be millions of specialized intelligence models.

>> Yeah, there's

>> different—explain general intelligence and specialized intelligence. People are not—

>> may not understand the difference.

>> Explain that. General intelligence versus specialized intelligence.

>> Yeah.

>> So you can think about general intelligence as the process of learning from common tasks that we humans perform.

>> So for example, analogy is when we were born, All of us have IQ that doesn't change, right?

>> So we human beings species, our IQ is way higher than the rest of the planet species. So that is when a generalized intelligence model was born, it carries the fundamental IQ that solves a lot of tasks, common tasks, because it learns from public internet where, by volume it knows how to solve this problem, that problem, and so on. So that's a general intelligence model, and you just use it off the shelf as is. But specialized intelligence are domain-specific problems or product-specific problems that only the product designer, they will create that. There's no way a general intelligence model will ever be able to learn from it because it has tons of in-product context to optimize it. For example, we— my background is we worked many, many years at Meta. One type of AI problem is recommendation, right?

>> But recommendation is personal.

>> Every single person will have their own preferences based on their interests, based on their background.

>> Based on their region. And there's no uniform way to predict what you like, what I like. And we all want to see different things, different video recommended, different product recommended, different newsfeed recommended. So this is just one example.

>> Obviously, there's a ton of other examples. So this type of personalized recommendation is just one form of specialized intelligence. But there are many other forms of specialized—

>> people can rationalize, okay, use OpenAI, ChatGPT, Claude. Hey, I want to do a task. They give it a prompt, they get an answer. Maybe it reasons, multi-step reasoning.

>> Okay, great.

>> But the enterprise, OpenAI doesn't crawl that. So again, it's locked.

>> Yes.

>> So the unlock of the private data becomes a big thing that's specialized.

>> Right.

>> And workflows are also not applications. Applications are like, hey, I built an app. And you ship it, put it in the App Store. Workflows are how companies run their business. And that has data behind it. And it's behind the firewall and it's known to them. That is what I think specialized is.

>> That's specialized. And the other way, a form of very popular specialized intelligence is the orchestration of your workflow and the orchestration of how you call into different AI models.. And that orchestration today is called harness.

>> It's specialized, it's customized, bespoke. And that harness usually is also powered by a model behind, and that logic is specialized towards this particular application. And who designed this? There's no common knowledge of how to orchestrate this. So those are all examples where you, the product design or a specific implementation design, need to co-train with the model so model really learn what you want to optimize for. So there are gazillion number of those examples where just a general purpose is not good enough and you need to have specialized. General purpose doesn't give you control.

>> Yeah.

>> Specialized gives you full control based on open weights. That's where the significance of open-weight model is here.

>> And you have this proprietary intelligence powering your product, make your product better. Your product drives more data, make the model better, better model, better product.

>> And then you start to play this.

>> Lin, you mentioned harness. I love that term. It comes up a lot. Harness is a mechanism for managing agents and tasks, let agents do their thing. When you think about open-source software, we love open source. We use all that. It's open-source software. You take it and you modify it.

>> And that works.

>> That's been generations of success have been built on open source. So what you're saying is open models, you take it and you modify it. This is why OpenWeights is so important because people don't understand that if you can use the OpenWeights, you can tailor it to your needs. Like open source software. I can get open source software.

>> I can add to it.

>> Now, the difference is you can keep the weights inside the company. This is a huge enterprise value proposition. So how are people using open models? What's your experience? You got a lot of customers.

>> You guys are the leader.

>> How are people setting up the harnesses? What are they doing with the open model? Because it makes total sense. I got data, I got engineers, I take the model. How do— how are people taking the open models and tweaking them?

>> Yeah. So first of all, I think end of last year we see a huge uptick of leveraging open models, leveraging open models as is, because again, the quality has passed a threshold.

>> And second is significant interest in customized open model to turn that into their own model, not one per company, one per task across a wide variety of tasks. So in a way, actually, the whole entire, I think, industry starts to learn quickly how to manage open models.

>> Exactly as you mentioned, the benefits are full control.

>> You can change all the weights, but it has a really solid baseline for you to build a delta on top of.

>> It's so much easier to tune.

>> And second is it's much cheaper.

>> So the open models, the cost can— you can bring down the cost.

>> Everyone's doing token maxing, right?

>> And then they look at the— they're carefully looking at the bill.

>> It's like, what is ROI here? Can you justify it? No one can. So you can bring down the cost without losing quality by 5 to 10 times. Massive, significant cost reduction.

>> And these two values combined are creating a huge explosive adoption of open models.

>> So the way to do that is first you need to define a judgment of whether your tuned model, how good they are.

>> So this is called eval. People start to create evals and evals are very familiar concept in general software development because we all need to write unit test, integration test, right?

>> So there are a lot of kinds of testing, and it's just a form of testing quality. And it is not different.

>> It's like software development.

>> Yeah.

>> And obviously, these tests need to continue to be more comprehensive.

>> But the key point is, once you have these evals, and then everyone has data. And then we can help you hill climb. Hill climb means you just optimize for those evals. And it's the more we have so many cases again, we can get there. And many leaders already start to pioneer this even from 2 years ago. Now, Cursor is one of our flagship customers and they just launched a new model yesterday, actually. Yes, Grok 4.5.

>> So they have been leveraging their data to tune the model.

>> Specially designed for the coding use cases.

>> Really, really good. And we have been partnering with Harvey, for example.

>> They establish a legal bench that is their eval, and we help them climb towards that benchmark.

>> There are many, many other examples. I think this year we also see— last year was the coding year.

>> This year we see a lot of the co-worker applications, there are general purpose co-workers, there are specialized co-workers across healthcare, insurance, legal, recruiting, marketing, sales, customer support across the board.

>> Very interesting.

>> And they all have their internal evals because if you want to really be a company, eval is the first step. Without eval, you cannot really do hill climbing.

>> Hill climbing.

>> Yes. What's the next progression?

>> And hill climbing is actually we provide a variety of tools.

>> With all different kinds of algorithms. Obviously, the most well-known one is actually reinforcement learning.

>> The interesting thing is for people to do reinforcement learning, our philosophy is we anchor on the best model. So we do tuning on the best, highest quality model to hill climb there. And then once,

>> because that's—Explain what hill climbing is first. I love

>> this.Yes. So hill climbing is literally like you climbing a mountain, right? So basically you say,

>> hey, I—Get in shape.

>> I want to climb that

>> mountain.That's the peak I want to get to.

>> The peak is our eval. And here, it's a training process.

>> And the training process basically says, this is my objective. And the training is going backwards, forward, backwards, forward in order to optimize for that objective and then get there.

>> And sometimes it cannot get there, it collapses.

>> Then, this

>> fails.Fall down, you get back up.

>> Yeah.You get back up and you change a bunch of things and try again.

>> So that's the analogy we use a lot. So after you have done that, and then you will get a good model, and then we use—

>> we immediately optimize for inference. So our platform is training inference co-design, co-optimized.

>> So because at the end, training is not the endpoint of this process, it really needs to deploy to inference, let your product use it. And then you do—

>> a lot of people do A/B testing because they compare A and B with or without this new model and see whether people like the product change or not.

>> Right.

>> So this is part of experimentation. Once they see strong product feedback, then they scale it.

>> So we are just one platform for you to do everything together.

>> And our job is to make it super easy that everyone can have specialized intelligence model.

>> And our vision is the future.

>> There will be millions of them.

>> And people are like, this is way too big.

>> But we really believe that. We really believe every single app company, every single enterprise, they deserve to have their own specialized model.

>> It can be scary for folks to see the big picture. I want to get to what your platform does because I think this is super important. You have two schools of thought. I want to go fast or I want to build my own.

>> Yes.

>> And you guys work with Harvey, you mentioned Cursor, you have a lot of the marquee AI companies. What's the reason why they use Fireworks? Is it because they want to go faster? You guys are providing the tooling, right? the skeptics say, well, I don't need Fireworks, I just do it myself. Explain the value, because I think there's a lot of tooling that you guys are providing.

>> There's evidence.

>> Yes.

>> Talk about that value, but why Fireworks?

>> Yes.

>> What's in it for me? What's the benefit? Right.

>> So I had an interesting conversation with Jensen after his GTC keynote. He mentioned one thing. There's no such thing as a general company. There's no such thing as a general company.

>> Every company exists for a reason because they add exceptional value in the ecosystem and they specialize really well. On a certain dimension.

>> So this is a product design principle.

>> Same for intelligence.

>> In our view, product and intelligence today, they are not decoupled.

>> They are one thing. So every company needs specialized intelligence. So that's our view.

>> And that's where the industry is going towards because, as I mentioned, it's existential

>> It's faster with you guysIf I want to do my specialized intelligence, I'd go with Fireworks. I'm on board. You're optimizing for that goal.

>> Right.

>> So I'm behind from day one. The catch-up, even then, I'm not optimized.

>> Right.

>> So we want to— we feel the product team or enterprise, they own their data. They own the brain. And we are the hands. For them to implement what the brain wants. So that's how we think about how we help them accelerate this journey. And in order for the brain to operate, to build this intelligence, they need to have deep AI knowledge.

>> They also need to have deep AI system knowledge to manage the whole entire thing.

>> And there's a lot to build. And the resource pool, a talent pool, is very scarce.

>> And we're like, hey, we can build this layer.

>> You build on us.

>> You focus on your biggest differentiation and we help you there.

>> So that's how you think about it.

>> You help them scale too. We have scale as a huge piece.

>> Absolutely.

>> Yes.

>> What's your biggest focus this year as you guys continue to have the momentum? By the way, I love the product.

>> What do you guys do?

>> You get people up and running and you continually add value.

>> To your point, what's your focus this year?

>> So our focus this year is to make our product way easier to use. Hey, we're living in the age of agents. We are in the process of automating a lot of things so our customer can use our product and run really fast. The second is to reach out to many people who are in need of—

>> we see a lot of demand, so we just need to grow our company and scale it so a lot of people can access our product. So those are the key focus, but we really believe we are on the right track of deploying this specialized intelligence to everyone. But as we talk about that, there's a huge debate across the industry about, hey, should there be one company owning all the intelligence or it should be everywhere democratized?

>> I think AI is such an important technology, piece of technology.

>> It shouldn't be owned by only one company.

>> It just feels— doesn't feel right.

>> You're enabling people to turn on their corporate brains.

>> Yes.

>> Basically.

>> hey, and you're not threatening anyone because you're open.

>> Yes.

>> And you're scaling people faster.

>> Absolutely.

>> Yes.

>> when ChatGPT came out, Dave Vellante and I published a very controversial chart called the Power Law of AI Models. It was basically a curve. It had a torso and a neck. And we speculated that we think there's going to be specialty models.

>> And guess what?

>> You're now pioneering the new frontier. And if you look at the data in the enterprise, only about 1% of their data is unlocked. That means that every company is sitting on a treasure trove of data that could be activated and turned on and connected into a global brain.

>> That's what you're doing.

>> That's the beauty of the time. the opportunity is so big here.

>> So that's why we are super excited to help everyone. Get there.

>> Well, Lin, great to see you in person here in theCUBE in Paris.

>> We'll see you in the Bay Area. Thanks for coming on theCUBE, really appreciate it. Congratulations on all the success and keep going.

>> Thanks for having me.

>> All right, the specialized model, specialized intelligence, this is going to unlock the enterprise. Coding was step one. People can see the value of coding, code generation that opens up the value is undeniable when they start seeing the actual data of the enterprise being codified and enabled for AI applications, the brain gets turned on, AI works, the benefits are coming. And again, as the AI infrastructure continues to build, more and more of the enterprise stack will see this capability. We're here at the RAISE Summit doing our part to bring the action to you. Thanks for watching.