Base44, the vibe coding platform that Wix acquired for $80 million just a year ago, was just six months old and had a team of eight people. We started rolling out our own AI models to help users create apps in natural language.
The move comes amid a growing debate in AI circles about whether frontier models are the best fit for all use cases. A related question is whether a business built on someone else’s model can really be defended in the long term. Bay Area-based Base44’s latest move speaks to both.
Although the custom LLM is still in its early days of deployment, Base44 hopes it will eventually outperform the Frontier model. According to founder Maor Shlomo, “Training and owning the model as part of[our]overall stack allows us to further optimize latency, cost, and efficiency.”
At first glance, this could be a way to stay ahead of competitors such as Lovable, a Swedish startup that achieved unicorn status with its Series A round last summer and relies on external LLMs. But Shlomo expects other players to train their own models. “At least a player that has gotten big enough and fast enough to get enough data.”
Jonathan Userovici, general partner at venture capital firm Headline, said his company’s portfolio includes AI companies such as Mistral AI, but not Base44. Data is one of the three key elements of defense for AI startups, along with distribution and technology stack.
As a result, players with strong brands are relying on data and infrastructure to improve their defenses, and Base44 fits into that pattern. The company says the first iteration of LLM, Base1, was developed and trained on datasets generated from “tens of millions of real user interactions on the platform.”
This dataset will continue to grow with your company. But so will his rivals. ” The bigger competitors may not be Vibecoding startups, but participants from Frontier AI Institute, which is moving closer to Base44’s home base—Cursor and Grok’s parent company xAI are both now part of SpaceX, and Claude Code is a Vibecoding player in its own right.
This gives Anthropic and other foundational AI providers access to data and feedback loops that can be used to improve models for app creation, but Shlomo believes that specialization will give Base44 an edge. “Models are advancing, but what they can do will remain very general,” he predicted.
Userovici cautioned against underestimating frontier models, citing the example of legal tech startup Harvey, which abandoned plans to train its own models. He doesn’t expect applied AI companies to become Frontier Labs en masse, but says he frames the Base44 move in a broader context, one in which inference costs are a key part of the equation.
This cost pressure is driving the changes that business customers are now demanding, Mr. Uselovich said. “Having the latest model for every use case doesn’t necessarily give you[return on investment]. That’s why the entire infrastructure is set up to orchestrate and optimize to choose the right model so that costs don’t go up while maintaining the same or similar performance for the majority of use cases.”
Although enterprises are still a minority among users of the Vibe coding platform, they account for a growing share of platform revenue, and users of all sizes are beginning to express concerns about the cost of using AI. Base44’s decision to develop its own LLM was driven by multiple factors, with cost savings likely being one of the benefits.
“What we want is a model that is more in line with what we think is right, more optimized for what users like in terms of the results they get, and ultimately faster and cheaper for our customers than using a state-of-the-art model like Opus,” Shlomo said.
When it comes to Base44 itself, the cost savings are less obvious. “Model ownership gives Base44 direct control over compute and inference spending, which is expected to structurally strengthen margins over time,” the company explained in a press release.
Even if profits lag, improved margins would be good news for Base44’s parent company, which recently announced it would lay off 20% of its employees. By contrast, Base44 has grown in employee numbers since the acquisition and announced a few months ago that annual recurring revenue exceeded $100 million.
This is still lower than Loveable, which announced earlier this month that ARR hit $500 million. But Shlomo is betting that the “huge engineering effort” to develop Base1 will solidify Base44’s position as “the only vertically integrated vibecoding application, or in Userovic’s words, the player that owns the distribution, the data, and the infrastructure all at once.”
If you buy through links in our articles, we may earn a small commission. This does not affect editorial independence.
