Base44 Trains Its Own Model as Vibe Coding Hunts Defensibility

Picture an Italian restaurant that buys the pasta wholesale for years, then one morning the owner walks in with a flour mill in the back of a van. That's roughly what Base44 just did to its supply chain. The Tel Aviv vibe coding outfit has stopped renting frontier intelligence by the token and started milling its own.

The model is called Base1, and the bet behind it is simple: when you own the mill, you control the margin. Whether that bet pays off is the more interesting question for everyone building on top of OpenAI, Anthropic, or Google right now.

What Happened

Base44, which Wix scooped up roughly a year ago for $80 million when the company was barely six months old and running with a team of eight, has started rolling out its own large language model. According to TechCrunch, the model is named Base1 and was trained on a dataset generated from tens of millions of real user interactions on the platform.

Founder Maor Shlomo is positioning this as a vertical integration play. He told reporters that "training and owning the model as part of [our] entire stack allows us a lot more optimizations on latency, cost, and efficiency." The company now describes itself as the "only vertically integrated vibe-coding application."

The financial backdrop matters here. Base44 passed $100 million in ARR a few months back, and headcount has been climbing since the Wix acquisition. That growth sits in awkward contrast to the parent company, which recently announced it would cut 20% of its workforce. Base44 is the asset that's working, and Wix needs it to keep working harder.

The competitor everyone points to is Lovable, the Swedish startup that hit unicorn status in its Series A last summer and crossed $500 million in ARR earlier this month while still relying on external LLMs. Lovable is bigger. Base44 is going deeper. Those are two genuinely different strategies, and the next twelve months will tell us which one ages better.

Shlomo also expects others to follow, at least the players with enough scale and velocity to have accumulated meaningful training data. So Base1 is less a flag in the ground and more the opening move in a new phase of this category.

Technical Anatomy

The guts of it: when you build on Claude Opus or GPT-class models, every user prompt is a metered API call. Margins get eaten by inference costs you don't control, latency you can't tune, and a roadmap dictated by whichever frontier lab you've bet on. Anyone who has watched their inference bill climb faster than their MRR knows the feeling.

Owning Base1 changes the unit economics in three places. First, inference moves from variable third-party cost to controllable internal compute. Base44's own press release language was telling: "ownership of the model gives Base44 direct control over compute and inference spend, expected to result in a structurally stronger margin profile over time." Note the "over time". This isn't a day-one win.

Second, the training data is the moat. Tens of millions of real user interactions on a vibe coding platform is exactly the kind of narrow, high-signal dataset that a specialist model can exploit and a general model can't easily replicate. Frontier labs like Anthropic, whose Claude tooling has itself become a vibe coding player, have breadth but not the same density of "user wanted X, got Y, edited it to Z" feedback loops on app generation specifically.

Third, latency and routing. Once you control the model, you can do the boring but valuable work: quantize for speed, fine-tune for your output formats, route easy prompts to Base1 and hard ones to Opus. Jonathan Userovici of Headline put this well: enterprise customers are now building "an entire infrastructure" for orchestration and optimization so that "costs don't skyrocket while maintaining the same or similar performance across the majority of use cases."

The part where it all falls over is training cost and talent. Shlomo himself called Base1 a "huge engineering effort," which is honest. A team that was eight people a year ago is now doing applied ML research. That's the bit that doesn't show up on the press release.

Who Gets Burned

The most exposed group: applied AI startups whose only differentiation is a clever prompt template wrapped around someone else's model. If Base44 is right that scale-plus-velocity equals enough data to train a specialist model, then everyone below that threshold is in trouble. Lovable, with five times the ARR, has the data. The thousand smaller vibe coding clones do not.

Frontier labs are also feeling the squeeze, though from the other direction. Cursor and xAI now both sit inside SpaceX. Claude Code has gone from API to product. The labs are moving down the stack into application territory at the same moment application companies are moving up the stack into model territory. They'll meet in the middle, and the middle is going to be crowded.

Userovici flagged the cautionary tale: legal tech startup Harvey reportedly abandoned plans to train its own model. Training is brutal. Not every applied company that announces "we're building our own model" will ship one that beats GPT-class output on their use case. Some will quietly fold the effort back into orchestration of third-party models and hope nobody notices.

Wix is the interesting case. A company laying off 20% of its workforce needs its acquired growth engine to deliver structurally better margins, not just topline. The board is almost certainly looking at Base1 as the lever that justifies the $80 million cheque retroactively. The pressure on Shlomo's team to translate Base1 into gross margin improvement, on a deadline, will be intense.

Enterprise buyers, finally, are the quiet winners. They're the minority of vibe coding platform users but a growing share of revenue, and they're the ones demanding the cost-control infrastructure that's now being built on their behalf. They get cheaper inference and better routing without having to build it themselves.

Playbook for AI Development

For platform leads and CTOs watching this play out, a few practical moves for the next quarter.

Start instrumenting your inference spend per user cohort yesterday. You cannot make a build-versus-buy argument on a custom model unless you know exactly where your token spend concentrates. If 80% of your inference cost comes from 20% of prompt patterns, that 20% is where a small specialist model earns its keep.

Treat your user interaction data as a balance sheet asset, not a logging artefact. The reason Base44 could even contemplate Base1 is because they were capturing structured outcomes at scale. Get your data contracts, retention policies, and consent flows in order now. Models trained on permissioned, high-signal data will beat models trained on scraped junk on narrow tasks.

Don't conflate "train our own model" with "abandon the frontier labs." Userovici's point about orchestration is the right read. The winning architecture is probably a small specialist model handling 70% of requests cheaply, with calls out to frontier models for the hard 30%. Build the router first. The custom model can come later, or never, depending on what the data says.

For open weights experimentation, the Hugging Face ecosystem remains the cheapest place to prototype fine-tuning approaches before committing to a full training run. Most teams should be testing LoRA adapters on open models long before they consider a Base1-scale effort.

Key Takeaways

• Base44 has launched Base1, an in-house LLM trained on tens of millions of platform interactions, positioning itself as the only vertically integrated vibe coding application.
• The economic case rests on inference cost control and margin expansion over time, not immediate savings; Wix needs this to work given its 20% workforce cut.
• Lovable's $500 million ARR on external LLMs proves the rent-a-model strategy still scales, so Base44's vertical bet is a contrarian one, not a settled answer.
• Frontier labs moving down the stack (Claude Code, xAI, Cursor) and applied companies moving up will collide in the middle of the application layer within twelve months.
• Most teams should build orchestration and routing before they build models; Harvey's abandoned training effort is the cautionary version of this story.

Back to the flour mill. The Italian restaurant owner with his own mill either makes the best pasta in the city or quietly sells the mill on Gumtree two years later and goes back to the wholesaler. Base44 has bought the mill. Now they have to prove the pasta is actually better, and that the customers can taste the difference at a price they're willing to pay. The vibe coding category just got a lot more interesting to watch.