Looking at the traffic of AI video platforms, and I'm a bit blown away — see that blue line that in late 2025 just shot straight up, leaving Runway, HeyGen, and Luma behind?
That's Higgsfield AI. And the team just officially became a "unicorn" (valued at $1.3B), closing an $80M round (total raised ~$138M). This time, in addition to early backers Menlo and DVC, the Accel team joined the round.
This is essentially the first Kazakh unicorn with a strong tech background. How did they outpace the market and hit $200M run-rate faster than anyone in history?
It actually all started with limited money. Kolya Davydov from DVC wrote a great story on his LinkedIn about how this project came to be:
In 2023, he introduced Alex Mashrabov (who had previously sold his startup AI Factory to Snapchat for $166M and got bored) to Yerzat Dulat — an RL genius from Kazakhstan. Yerzat was already a local legend because he trained the first Kazakh LLM for just $350k. When Kolya asked how he managed to do it so cheaply (usually millions are burned on this), Yerzat answered in true hacker style: "Well, I just didn't have more money."
And so it all began. But the turning point came when they pivoted to SMB marketing. While other projects were competing on evals and fluid physics, Higgsfield embedded slick growth hacks right into their code.
They realized that pleasing filmmakers is hard, while millions of sellers on Amazon/Shopify need "lots of videos right now." They understood that 99% of the money comes not from filmmakers but from SMB marketers.
They rolled out the Click-To-Ad feature — you drop a link to a Shopify product and the neural network automatically parses images, prices, descriptions, and assembles 10 ready-made ad clips. No prompts like "woman eating salad," just one click — and you've got creatives for A/B testing.
And betting on mobile (the Diffuse app) gave them a viral boost while competitors were sitting in Discord.
Higgsfield has its own Reasoning Engine and motion control models (based on Yerzat's RL), but for textures they can use other models (like Kling) if those perform better. Their main strength is control. While other neural networks produce random results, Higgsfield gives you a director's toolset: "I want this exact zoom" and "the face shouldn't change."
The scoreboard: in November they broke $1M in daily revenue. Congratulations!