Base44 launches Base1 to improve UI generation

In a June 29, 2026 blog post, Base44 founder and CEO Maor Shlomo announced Base1, a model trained specifically for building web applications inside Base44's platform. Shlomo told Calcalist (CTech) the model is fine-tuned on top of an existing open-source foundation model rather than trained from scratch, arguing a specialized model optimized narrowly for one task can outperform a general-purpose frontier model on that task while being faster and cheaper to run. Wix, which acquired Base44 for at least $80 million in 2025, also published its own post on the launch.

Per Shlomo, Base1 is trained on data from Base44's own building sessions, tens of millions of real user interactions, including what the AI built, what broke, and what users accepted or rejected, plus reinforcement learning in simulated environments, to optimize for both working code and better product decisions rather than syntax alone. Base1 joins GPT-5.5 and Claude's Opus 4.8 in Base44's model selector; Base44 previously depended entirely on external providers including OpenAI and Anthropic.

TechCrunch frames the move within a broader pattern of AI products with sufficient scale and data considering in-house models for cost control and defensibility, quoting Headline partner Jonathan Userovici on data, distribution, and tech stack as the three pillars of AI-startup moats. Base44 says it has grown to 2 million users and a $150 million annual recurring revenue run rate (up from $50 million as of last November) and calls itself the largest AI-powered app-creation platform in North America; competitors in the vibe-coding category include Lovable, Replit, and Cursor, which face the same build-vs-buy choice on models.

Base44's rapid growth has come with real security incidents relevant to any team shipping AI-generated code: Wiz previously disclosed a permissions flaw in Base44-built apps that exposed personally identifiable information and trade secrets, and Imperva separately identified critical vulnerabilities that could have let attackers access data or take over applications. Shlomo told Calcalist the platform now automatically scans every generated app for exposure and misconfiguration issues and plans to announce cybersecurity partnerships in the coming weeks.

The core trade-off for teams building UI- or design-sensitive AI products: owning a fine-tuned model adds real engineering overhead (data pipelines, RL loops, inference deployment) but can pay off in cost, latency, and output variety when a platform has enough in-product usage data to train on. Base44's security track record is also a useful cautionary data point for any vibe-coding or AI-code-generation tool: platform-level automated scanning is emerging as a baseline expectation, not a nice-to-have.

Shlomo has said Base1's first versions are meant to match, not yet beat, frontier models on app-building quality, with the ambition to eventually outperform them; no side-by-side benchmark against GPT-5.5 or Opus 4.8 has been published yet. Base44's promised cybersecurity partnership announcements are also worth tracking given its vulnerability history.