Base44, the vibe-coding platform (a method of building apps through natural-language instructions) acquired by Wix, has begun rolling out its own in-house AI model, "Base1." The move shifts the company away from the general-purpose, third-party large language models (LLMs) it had relied on, toward a specialized model trained on the vast trove of usage data it has accumulated. By optimizing response speed, cost, and the quality of the apps it generates for its own specific use case, Base44 aims to shore up its competitive edge as an AI startup.

A Specialized Model "Base1" Trained on Its Own Data

Base44 is a startup based in Tel Aviv, Israel, acquired by Wix about a year ago for 80 million USD (about 13 billion yen). At the time of the acquisition, it was only about six months old and had a team of just eight people. From there it grew rapidly, and it has now reached the point of owning its own model.

Base1 is a model built on an existing open-source LLM and fine-tuned specifically for the task of web app development. According to the company, it was trained on a dataset created from tens of millions of real user interactions carried out on the platform. The approach centers on reinforcement learning: the model is repeatedly put to work on real app-building and editing tasks, its outputs are judged as good or bad, and that signal is fed back into the model's weights to keep refining it.

Founder Maor Shlomo explains that training and owning the model as part of the entire technology stack makes it far easier to pursue aggressive optimizations in speed, cost, and efficiency. He believes general-purpose models will keep advancing but will remain oriented toward broad, all-purpose use, and he sees the real opportunity in a specialized model built for a narrow use case. Ultimately, the company wants to deliver results to users faster and more cheaply than using a frontier model (a cutting-edge, massive model) such as Opus.

The Key Is "Defensibility": Fencing In Data, Distribution, and Infrastructure

Behind this decision lies an intensifying debate over the defensibility of AI startups: the long-term worry that a business simply riding on top of someone else's model could be easily overtaken by competitors.

Jonathan Userovici, a general partner at the venture capital firm Headline, points to three ingredients that let an AI startup shore up its defenses: data, distribution (the path for reaching users), and the technology stack. His view is that companies with strong brands are increasingly digging into their own data and infrastructure to build something rivals cannot easily copy, and Base44's push into its own model sits squarely within that trend. Indeed, Shlomo says he wants to make Base44 the "only vertically integrated vibe-coding company," one that owns its distribution, data, and infrastructure all at once.

That said, the competition is not limited to other vibe-coding startups. Alongside direct rivals such as Sweden's Lovable, which relies on external LLMs, the frontier research labs themselves are drawing closer. Cursor and xAI, the parent of Grok, have both come under SpaceX, and Anthropic's Claude Code has grown into a vibe-coding player in its own right. These are formidable players that command huge volumes of usage data and virtuous feedback loops.

At the same time, there is a note of caution that going the in-house-model route is no cure-all. Userovici cites Harvey, a legal-tech company that abandoned plans to train its own model, and warns against underestimating frontier models. He does not expect applied-AI companies to turn into research labs en masse; rather, he frames the move as being driven by the growing weight of inference costs (the processing required to run a model and produce an answer).

Base44 Keeps Growing, With Hopes for a Better Cost Structure

Base44's business is doing well. The company announced that its annual recurring revenue (ARR) surpassed 150 million USD (about 24 billion yen) in May, just two months after it crossed 100 million USD (about 16 billion yen). That still trails rival Lovable, which said it reached 500 million USD (about 80 billion yen) in ARR this month, but Base44 has kept growing while adding headcount even after the acquisition. It stands in sharp contrast to parent company Wix, which recently announced it would cut 20 percent of its workforce.

*Converted at 1 USD = 162 JPY (as of July 2, 2026)

The cost benefits will not show up in the numbers right away. Even so, the company expects that owning its own model will give it direct control over spending on compute and inference, allowing it to strengthen the very structure of its margins over time. Developing Base1 is said to have required a substantial engineering effort, but Shlomo sees that investment as the foundation for cementing its vertically integrated position.

Summary

Base44's in-house model "Base1" is a bid to move beyond a business that merely rides on general-purpose external models. By training a specialized model on its own data and honing speed, cost, and quality for its specific use case, the company is trying to build an edge that rivals will find hard to follow. As frontier models keep advancing and inference costs grow heavier, this could become a bellwether for how far applied-AI companies will go in building their own models.