Alex Mashrabov, CEO of Higgsfield, on orchestrating AI video models

We first covered Higgsfield at $100M ARR in November 2025, up 9x from $11M in May, finding product-market fit with marketing agencies and e-commerce brands generating video ads.

To learn more, we reached out to Alex Mashrabov, founder and CEO of Higgsfield.

Key points from our conversation via Sacra AI:

What is Higgsfield in short and what was Higgsfield's initial product-market fit?

The path to build Higgsfield was not just one shot. We had several iterations to get there. We had a strong feeling that there is a new digital camera which should exist for social media. This starts from my time at Snap, where I saw the power of video—that video is the language of the Internet—and AR face effects and AR ads all opened new storytelling opportunities. At Higgsfield, we completely removed the production tax so that the best idea and best story should win.

We took a very different approach last year. We launched a consumer mobile app and scaled to 1 million users in two months, but retention was challenging. We realized that the true power users are typically VFX artists and professional marketers, and they need way more control, which is really difficult to achieve on mobile. We decided to focus on desktop products. Since we launched this on March 31st, it completely took off.

For us, the product market fit process is a challenge we have to address every day as the landscape changes every week. There is a major model release every week, so we have to constantly monitor social media. Since we release six days a week—more than 300 times a year—we have this closed feedback loop so we learn quickly. We're going through the early stages of adoption of video generation AI. We are scaling to 100 million professional creators who eventually will be using video generation AI.

Text generation had its "ChatGPT moment" when it took off, but video generation doesn’t seem to have had quite the same kind of big moment yet, even from Sora. Is that how you see it, and why?

Video is the most complex domain. If you think about it, humans can read maybe 200 words a minute, although one minute of video takes approximately 60,000 words to describe—it's the equivalent because video is way more dense. The same applies to these video models. Video generation requires roughly 100 times more tokens than text. It is just a more complex domain. That's why it takes more time to develop the models.

Many companies and research labs in the world agree with us that we will unlock new reasoning capabilities and true world modeling by scaling these models. Some companies say this is actually a path to AGI because it unlocks that true world modeling.

Broadly, your customers are creators and marketers, but those categories are pretty generic. How would you describe your customer in your own words? Can you talk about why pre-Higgsfield products didn’t serve them well and what your customer had to do pre-Higgsfield to accomplish what Higgsfield now enables them to do?

Let's break it down by two categories. The first category are people who are pushing commercial work out, and the second are those using AI for brainstorming.

For the first category, we see tens of thousands of AI-first creators who make commercial marketing campaigns on the Higgsfield platform for many major brands in the world. Fundamentally, they see an immediate opportunity to cut down budgets from, let's say, $100,000 for one minute to maybe $500 per minute. It's 200 times higher cost efficiency for them. The only bottleneck for these AI-first creators is actually communication with the clients and defending and explaining their creative vision.

On the second front, there are the so-called incumbents. Think about existing agencies. Most of them already use Higgsfield for storyboarding, although there is still substantive resistance from putting AI-generated work in production. One brand is working on their Super Bowl ads using generative AI and our platform, although it's not going to be fully end-to-end AI-generated.

You recently launched the Higgsfield API. Who is the customer for it and how do you anticipate them using it? How do the economics and value proposition differ between these two surfaces? Is the API about serving models at scale with margin, or is there more to it?

Since it's a beta product, we have not announced this yet. We typically keep the metrics confidential, although I'm happy to share an exclusive data point: we're seeing customers with marketing budgets over $100 million who turn 90% of their ad creative—their social media ad creative—to be generated with AI. It is absolutely astonishing.

I've worked in the video AI space my whole professional career—the last ten years. It is truly astonishing to see that the pace of acceleration in 2025 is maybe 10x from what it was before. We made more progress over the last nine months than we did in the previous nine years. 2025 has become the year where many marketing teams throughout the world, and especially agencies, see extremely high efficiency by using image and video generation AI platforms.

Unlike marketplaces like Replicate, fal.ai, or Hugging Face, which offer vast model selection, Higgsfield seems to curate models, focusing on packaging them into use-case specific solutions for marketers and creatives. Is this targeted, use-case emphasis Higgsfield's core strategy, and why is this approach better than offering access to every available model?

That's a great question. We're fundamentally different. fal.ai is our partner. We are deeply grateful for the relationship we have. They're fundamentally a model aggregator where they deliver access to various closed-source models, and they're famous for their reliability for developers who build on top of those models. This is also great for experimentation as they have access to hundreds of models.

For Higgsfield, it's completely different. We orchestrate the entire workflow. Our goal is to deliver the video and eventually deliver sales through videos. To achieve that, we play at the model layer where we do post-training, fine-tuning, and auto-prompting, auto-selection of model for specific use cases, especially around social media. That's how we are different.

How does Higgsfield's business model compare to Photoroom, a more traditional SaaS for background removal and other workflow tools around images with monthly costs. Do you think of Higgsfield as a near-traditional SaaS or SaaS-like company?

This is a great question. SaaS is definitely the closest business model, but generative AI redefines growth expectations entirely. Specifically about Photoroom, it's a great product, and it shows us vividly how much Adobe misses multiple opportunities, as Photoroom covers maybe the same use case as Adobe Express. But you can see when the quality of even a SaaS product is great, how it can be well differentiated.

We think of ourselves as a cutting-edge generative AI company. Generative AI redefines growth expectations entirely. You know that in SaaS, they like to say that a YC company at the scale of maybe a couple million ARR should be growing 30% month over month, and this is the golden target. This is the way SaaS companies used to grow.

Now when you look at the fastest-growing companies like Higgsfield, Cursor, Lovable, all these companies grow way faster than that at way higher scale. For us, we believe we are just at the beginning of the adoption curve.

When we think about the leading companies in adjacent space, Canva comes to mind. Canva has maybe 20 to 25 million paying users. The adoption of video generation AI today is maybe 1 to 2 million paying users overall. This gives us a sense that the adoption will grow at least 20x from where we are today.

At the same time, we believe this is a net new market, so there are going to be second and third-order effects which we cannot project just yet as more and more content is created. We believe that eventually more than 50% of content on social media is going to be generated with AI.

Regarding Higgsfield, since you use your own fine-tuned and post-trained models, do you see it as a blend of a research and product company? This is similar to Runway—maybe more for professional filmmakers—and Kling, which now offers creative suite workflow tools. Is that an accurate view?

Absolutely. At the end of the day, the goal is to build the best product for social media creators, and we found that the best form factor is the desktop through a series of iterations. We also learned that a broader set of social media creators don't necessarily care about the specific model they use. They need a platform where they can get state-of-the-art AI capabilities, and that's what we strive to do at Higgsfield.

Many companies you mentioned struggle to sustain a competitive edge due to really long model development cycles, so they can iterate one quarter at most. Higgsfield can iterate every day. We believe this allows us to deliver the best product to the market.

Thanks to the very close internal feedback loop, we have AI researchers, prompt engineers, and professional VFX artists who used to make commercial ads for top brands. By having them all working together, we deliver the best product as we are the users of our own product.

Going back to the feedback cycle and learning from social media, what specifically are you looking for? Success stories of AI video, or something else?

This is actually a great question. We approach this completely differently. We strongly believe that the evolution for Higgsfield is to become a platform that supports marketers and creative marketers to achieve their goals and eventually deliver higher engagement and higher sales.

To do that, every marketer needs to constantly monitor trends, analyze the performance of existing videos, figure out new creative production, allocate budgets across the most performing videos, and then rinse and repeat this every day. We already do this internally for our own product growth.

This marketing automation product is already in beta, but we're also internally keeping it polished. At some point, we will launch this publicly as a new and hopefully state-of-the-art platform for marketing automation for social media.

Could you elaborate on that? Specifically, how would Higgsfield's existing generative video technology be integrated or combined to enable marketing automation?

I can give you an example. In a recent speech, Andrej Karpathy said that reinforcement learning is not a path to AGI. We tend to agree with him, although we believe that Higgsfield will be one of the winners of broader adoption of reinforcement learning as we can optimize the whole feedback loop in marketing, the whole process with reinforcement learning. We apply reinforcement learning for our own video models and image models training.

At the same time, we see a way bigger opportunity when we expand from video generation to the whole workflow. Today, Higgsfield covers the workflow from ideation to video creation and then collaboration, so the natural next steps are publishing and measurement.

Given the fast-changing model landscape, with new models appearing daily on Higgsfield, how do you integrate them? Do you test new models on existing workflows to determine if they offer improvements before integrating them?

Absolutely. We benefit from each new model release. Customers choose Higgsfield because of the workflows we provide, so we have to absorb the complexity. We have to become the experts for every new model. We have to be the best in the market to integrate each new model into a usable commercial workflow.

It is also important to mention that we take an approach which is different from many other research labs as we focus on use cases and actual adoption. As creator behavior adapts and the adoption scales maybe by 10 to 20 times from today's moment, we will see the whole trillion-dollar social media marketing space completely changed.

You have seen that already. Some companies like Adobe took a 45% stock price hit, which is $130 billion, as an indication of the importance of AI and the AI-first strategy. We believe that AI-first means to be the fastest to deliver state-of-the-art AI capabilities to professional social media creators.

How important is the open source ecosystem of models to Higgsfield? Does the open source vs closed source use case differ and if so, how?

We learned how to capitalize on both. Open-source models allow us to build a much more customized experience through post-training and fine-tuning. Also, it allows us to optimize, to distill the models, and really deliver better gross margin for specific use cases. Almost all closed-source model providers work with us closely so we help them optimize their models through post-training to deliver the best user experience.

You mentioned your research advantage comes from shorter iteration cycles compared to competitors' long ones. Is this because you focus on post-training and fine-tuning rather than building entirely proprietary models, or is there another reason for your faster cycles?

Absolutely. While today the focus of many companies and public attention has been really about the video models and their capabilities, we have entered the era of commercial video production using generative AI very quickly. We are essentially focused on Higgsfield becoming the go-to platform for commercial social media production with generative AI. This is fundamentally very different thinking from building the best model in the world.

What do you make of other sort of video AI trends such as video avatars or talking heads? Is that something that you would do e.g. if your customers asked for it?

We focus on high-end visual fidelity and also synthetic creators who can be synthetic spokespersons, who can become brand ambassadors. Most viewers will not be able to tell that these videos are actually generated with AI. We focus on media that feels real and evokes emotion.

What are your plans for your Almaty, Kazakhstan office and what’s its strategic importance to Higgsfield?

We're building a global company with headquarters in California, and we do have a strong presence in Central Asia. This gives us a unique cultural bandwidth as Kazakhstan has a mix of South Korean culture, Middle Eastern culture, and American culture. They also have a really cracked engineering team. Central Asian talent overall consistently ranks in the top 20 in the world across math, programming, and physics competitions.

Myself, I was top three in the world in programming competitions, and we have 11 such winners of international competitions on our team. Overall, when we look at the space, we have a deep appreciation for the potential of people from that region. Kazakhstan itself is home to over 10 consumer-focused unicorns. It provides us with a high concentration of talent that is hungry to keep building global products.

If everything goes right for Higgsfield over the next 5 years, what does it become and how is the world changed?

In five years, by the end of this decade, we expect that most of the content on social media is going to be generated with AI, and most of that AI-generated content is going to be created with Higgsfield.

This is a completely new market which no one touches, including Adobe and Canva today. We believe that if we follow the trajectory, if we follow the path, we can become a sustainable $100 billion company.

This transcript is for information purposes only and does not constitute advice of any type or trade recommendation and should not form the basis of any investment decision. Sacra accepts no liability for the transcript or for any errors, omissions or inaccuracies in respect of it. The views of the experts expressed in the transcript are those of the experts and they are not endorsed by, nor do they represent the opinion of Sacra. Sacra reserves all copyright, intellectual property rights in the transcript. Any modification, copying, displaying, distributing, transmitting, publishing, licensing, creating derivative works from, or selling any transcript is strictly prohibited.