Base44, the vibe-coding platform that Wix acquired for $80 million just a year ago, has started rolling out its own large language model, Base1, to help users build apps from natural language, according to TechCrunch. The first iteration was trained on a dataset generated from 'tens of millions of real user interactions' on the platform, and Base44 says it eventually hopes Base1 will outperform frontier models on its specific task.
The move is a deliberate answer to one of the loudest debates in AI: whether businesses built on top of someone else's frontier model are truly defensible. As model quality converges and costs fall, the worry is that application-layer startups are just thin wrappers whose value can be replicated or absorbed. Base44's bet is that owning the model changes that calculus -- the company framed it explicitly, saying model ownership gives it 'direct control over compute and inference spend, expected to result in a structurally stronger margin profile over time.'
“The move is a deliberate answer to one of the loudest debates in AI: whether businesses built on top of someone else's frontier model are truly defensible.”
The defensibility logic rests on data. As Headline GP Jonathan Userovici frames it, data is one of three key ingredients of durable advantage for AI startups, alongside distribution and tech stack. Base44's tens of millions of real building sessions are exactly the kind of proprietary, task-specific signal a generic frontier model cannot easily acquire -- and the wager is that a smaller model trained on that data can beat a larger general one on the narrow job of turning prompts into working apps.
The competitive backdrop is brutal and converging. The bigger threat to Base44 may not be other vibe-coding startups but the frontier labs themselves: Cursor and xAI now both sit under SpaceX, Claude Code has become a vibe-coding force in its own right, and -- as of this week -- Meituan's MIT-licensed LongCat-2.0 gives anyone a free 1.6T coding model to build on. In that landscape, owning a differentiated model trained on proprietary data is one of the few ways to avoid being commoditized from above.
The bear case is that training and serving a competitive model is expensive and hard, and a smaller in-house model may lag frontier systems on quality even if it wins on cost and fit. There is real execution risk in a company best known for app-building suddenly running its own AI training. What to watch: whether Base1 actually outperforms frontier models on Base44's task, how the margin story plays out, and whether other application-layer startups follow the build-your-own-model path.