Wix-Owned Base44 Builds Its Own AI Model, Betting Ownership Beats Renting Frontier Intelligence

Base44, the vibe coding platform Wix bought for $80 million when it was barely six months old, has begun rolling out a proprietary AI model of its own, a move that puts a concrete bet behind one of the more contested questions in AI right now: whether building on top of someone else's frontier model is actually a sustainable business.

From Eight Employees to Training Its Own LLM

The speed of this trajectory is worth sitting with for a second. Wix acquired Base44 roughly a year ago for $80 million, at a point when the Tel Aviv company had existed for six months and employed eight people. It was, by any normal startup timeline, still finding its footing. A year later, it's shipping a fine-tuned language model trained on its own user data, built in partnership with Wix's existing machine learning team, the same group that previously built Wix's Harmony model.

The new model is called Base1. Founder and chief executive Maor Shlomo told The New Stack the project started several months ago, and that even he wouldn't have predicted Base44 would be training its own model this soon. Two things changed his timeline: open-source base models matured faster than expected, giving the team a strong foundation to fine-tune rather than build from zero, and Base44's own growth gave it enough real usage data to make that fine-tuning worthwhile.

That distinction matters, because Base1 is not a frontier model trained from scratch. Shlomo was direct about this in his interview with The New Stack, noting that building a frontier model from the ground up requires several billion dollars, money no vibe coding startup has. Instead, Base44 took an existing open-source model and tuned it specifically for one job: building web applications inside Base44's own platform, using its own tooling and its own agentic harness.

How the Training Actually Works

The process leans heavily on reinforcement learning rather than static fine-tuning alone. Base44 runs the model repeatedly against real tasks pulled from its platform, building an app, editing one, fixing something a user flagged, and scores each output as good or bad. That signal feeds back into the model's weights. Most of the training data, according to Shlomo, comes from running the model on Base44 itself and learning from how it performs, layered on top of a smaller base of the tens of millions of real user interactions the platform has logged.

Shlomo isn't claiming Base1 immediately outperforms general-purpose frontier models. In comments to Calcalist, he was careful about expectations, saying the first versions will perform on par with the best available coding models, not necessarily better. What he's betting on is the trend line. Faster and cheaper now, with the hope that specialization compounds into a genuine performance edge over time as the model keeps learning from how actual users build on the platform. Base1 is also explicitly the first release in a planned series. Shlomo describes larger models and deeper product integration as already in motion.

The Defensibility Argument, and Its Skeptics

Step back from Base44 specifically and there's a broader industry argument playing out here. If every AI-powered app-building platform is wrapping the same handful of frontier models from OpenAI, Anthropic and a few others, what actually separates one from the next? That question has been circulating in venture circles for a while, and it's becoming sharper as the number of vibe coding platforms multiplies.

Jonathan Userovici, a general partner at Headline whose portfolio includes Mistral AI, frames defensibility for AI startups around three ingredients: data, distribution and tech stack. Base44 is explicitly trying to check all three boxes at once. It already owns distribution through its user base and its position inside Wix's broader product suite. It's accumulating data through tens of millions of real interactions. Owning the model is the piece that completes the stack.

Userovici isn't fully sold on the strategy applying everywhere, though. He pointed to Harvey, the legal tech startup that explored training its own model and ultimately abandoned the plan, as a reminder that going vertical doesn't automatically pay off. His read on Base44's move is less about chasing model supremacy and more about cost control. Inference spend has become a real line item for AI companies, and enterprise customers in particular are pushing back on paying premium frontier-model prices for every single task. An entire layer of orchestration tooling has sprung up just to route requests to whichever model is cheap enough for the job without sacrificing quality, and Userovici sees Base44's in-house model as one version of that same instinct, just brought entirely in-house.

Shlomo's own framing leans more philosophical than financial. He told TechCrunch that general models will keep improving but will stay broad by design, while a model built specifically for one narrow task, building web apps inside one specific environment, can develop a kind of product judgment a general model never will. He calls this a model with taste, in a company blog post explaining the rationale: not just a better coder, but a system that understands what separates a working app from a mediocre one and pushes back when a user's instinct points somewhere worse.

The Real Competition May Not Be Other Vibe Coding Startups

Base44 competes most directly with Lovable, the Swedish startup that hit unicorn status off the back of its Series A last summer and has stuck with external frontier models rather than building its own. Lovable is also, by revenue, well ahead of Base44 right now. It announced $500 million in annual recurring revenue earlier this month, more than triple Base44's own milestone of $150 million in ARR, which it crossed in May just two months after passing $100 million.

The more interesting threat, by Shlomo's own admission, isn't Lovable or any other vibe coding pure-play. It's the frontier labs themselves moving directly into app-building. Cursor and xAI now both sit under the SpaceX umbrella. Claude Code has evolved well past a coding assistant into something that competes directly for the same vibe coding use case Base44 was built around. That puts Anthropic and other foundational AI providers in an unusual position: they get the same kind of usage data and feedback loops Base44 is relying on to train Base1, except at a scale no single vertical platform can match.

Whether narrow specialization beats that scale advantage is exactly the bet Base44 is making, and it's a bet the company won't be able to fully prove out for a while yet.

Growth Comes With Baggage

Base44's growth story hasn't been entirely clean. Security researchers at Wiz disclosed a serious permissions flaw last year affecting applications built on the platform, one that exposed personally identifiable information and trade secrets belonging to thousands of organizations. Imperva separately identified critical vulnerabilities that could have let attackers access sensitive data or take control of apps built through Base44. Neither issue is unique to Base44 specifically. Security researchers have repeatedly flagged that AI-generated code tends to carry vulnerabilities that non-technical users, the exact audience vibe coding platforms are built for, have no way of catching on their own.

For Wix, the parent company, the timing of Base44's growth carries some irony. Wix itself recently announced it would cut 20% of its workforce, even as its AI-native subsidiary keeps adding headcount and revenue. Base44 passing $150 million in ARR and unlocking a reported $38 million payout for Shlomo stands in fairly sharp contrast to the financial picture at the parent level, where shares have fallen sharply this year despite the company's broader push into AI products. Whether Base1 meaningfully widens Base44's margins, as the company expects, or simply slows the rate at which inference costs eat into them, will likely become clearer over the next few product cycles rather than this one.