Base44's Bespoke AI Model Takes on Frontier Models for App Building

Base44, a Wix-owned platform for building web applications with AI, has unveiled Base One, its first proprietary language model tailored exclusively for converting user prompts into functional web apps. The move represents a shift away from relying on general-purpose frontier models like Claude and GPT, betting instead that specialization will outpace versatility for this specific task.

The fine-tuning strategy#

Base One isn't trained from scratch; rather, it's a fine-tuned adaptation of an existing open-source foundation model. The training methodology relies on reinforcement learning, where Base44 repeatedly runs the model against real platform tasks—building and editing applications—then scores the outputs and feeds those signals back to refine the model's parameters. The training data combines outputs generated from the platform itself with synthetic data, with the bulk coming from actual Base44 usage patterns.

According to CEO Maor Shlomo, the core insight driving this approach is that generalist frontier models, while powerful across many domains, lack the specialized optimization needed for the particular constraints and patterns of web application development. "For many use cases, it might not be the best choice," Shlomo explained, noting that a model aligned specifically to Base44's design philosophy and product decisions should deliver superior results in its narrow domain.

Business and control advantages#

Beyond performance, owning the model stack offers material cost benefits. Managing inference and compute spend in-house rather than through external vendors directly addresses one of the largest expense categories for AI-native businesses. This also reduces dependency on third-party API providers for a critical component of the product.

The acceleration toward this milestone surprised even Shlomo himself. A year earlier, he would not have predicted such rapid internal model development. Two converging factors changed the timeline: the maturation of open-source models provided a solid foundation for fine-tuning, and Base44's explosive growth—revenue jumped from $3 million to $150 million in under a year following Wix's $80 million acquisition—supplied the scale of data and traffic necessary to make fine-tuning effective.

Collaboration and future plans#

The development was a joint effort between Base44 and Wix's machine learning and data science teams, who previously built Wix Harmony. Base One is tuned to Base44's agentic framework, incorporating the platform's tools, instructions, and operational parameters.