Wix's Base44 builds its own AI model as race for specialized LLMs intensifies
Base1 replaces general-purpose models with an AI trained specifically for application development.
Base44, Wix's AI-powered application development platform, is launching its own large language model (LLM), replacing the general-purpose AI models the division has relied on until now.
"This is an important moment for the Israeli AI ecosystem, and we hope to see more companies making similar moves," Base44 CEO Maor Shlomo told Calcalist. "It is especially important at a time when the U.S. government is restricting access to some of the newest AI models. There are too few companies in Israel building and training models or taking on projects like this."
Base44 enables users to create applications and websites through natural-language conversations with AI, a process widely known as vibe coding. Until now, the platform has relied on foundation models developed by companies including OpenAI and Anthropic. It is now introducing Base1, an AI model developed and fine-tuned in-house on top of an open-source foundation model, with the goal of reducing dependence on the industry's largest AI providers.
"This is a process we started several months ago," Shlomo said. "Base1 is built on an existing open-source model. Building a frontier model from scratch requires several billion dollars. Open-source models are already incredibly powerful and are giving the largest AI labs real competition. We took one of those models and optimized it specifically for Base44's use case, building applications. That's our bet: a specialized model inside a platform will outperform a generic model while also being faster and more efficient."
According to Shlomo, specialization gives Base1 a significant advantage.
"The large AI labs release models that have to be good at everything, writing essays, generating software code, designing websites and answering general questions," he said. "We took a model and made it exceptionally good at building applications. It understands the Base44 platform and is better at delivering exactly what users need."
That specialization is possible because the model has been trained on the vast amount of data accumulated by Base44 over the past two years.
"We're sitting on enormous amounts of traffic and application data," Shlomo said. "There are already tens of millions of applications and an incredible number of users. We can see what makes applications successful, what causes failures, and why. In addition, the model is continuously trained in simulated environments. As a result, it's not only better at writing code, but also at making the right product decisions based on what users actually want to build."
Shlomo expects Base1's initial versions to match the industry's leading models for application development while offering lower costs and faster performance.
"In its first versions, Base1 will perform on par with the best models available for application development," he said. "But it will be cheaper, faster and, hopefully, have a better design sense. Over time, I hope it will outperform every other model in this category because it is trained specifically on the Base44 platform and understands our users better than anyone else. It will take time to get there, it's an engineering adventure."
Shlomo founded Base44 in late 2024, shortly after completing an extended period of reserve duty following the October 7 attacks. The startup was among the first companies to recognize the potential of large language models to enable non-programmers to build software simply by describing what they wanted in natural language, even before "vibe coding" became one of the defining trends of the AI industry.
Within months, the company had attracted more than 100,000 users and signed partnerships with companies including eToro and Similarweb. Just six months after it was founded, with only six employees and no outside funding, Base44 was acquired by Wix for at least $80 million.
Last November, Wix reported that Base44 had surpassed 2 million users. The platform was also adding more than 1,000 paying subscribers a day and had reached an annual recurring revenue (ARR) run rate of $50 million. By May, that figure had climbed to $150 million. According to the company, Base44 is now the largest AI-powered application creation platform in North America.
The platform's rapid growth has also exposed it to the security challenges facing the broader vibe-coding industry.
Among other incidents, Wiz disclosed a serious permissions flaw that affected applications built on the platform and exposed personally identifiable information and trade secrets belonging to thousands of organizations. Cybersecurity company Imperva also identified several critical vulnerabilities that could have allowed attackers to access sensitive information and take control of applications.
More broadly, security researchers have repeatedly warned that AI-generated code can introduce vulnerabilities that inexperienced users may fail to detect. While experienced software developers often identify such issues before deployment, users with little or no programming knowledge may not recognize them at all.
According to Shlomo, Base44 has invested heavily in addressing that problem.
"Today, our platform includes multiple layers of security," he said. "Every application is scanned automatically for issues such as data exposure, insecure code, and configuration mistakes. Organizations can also define strict permissions and control who can use each application. We have partnerships with major cybersecurity companies that we'll announce in the coming weeks."
"The biggest challenge," he added, "is making security accessible to non-technical users. We want to make sure that what people publish is exactly what they intended to publish, and nothing that could be exploited by attackers. Every version of an application is automatically scanned for vulnerabilities. If the AI builds something insecurely, exposes information, or creates a dangerous configuration, the user is alerted before deployment. It's a very robust security layer."
