Base44 Launches Custom LLM to Defend Against Frontier Model Competition

Base44’s Defensive Play in a Crowded Market

Base44, the Bay Area no-code platform acquired by Wix for $80 million in 2025 when the startup was just six months old with an eight-person team, has begun rolling out Base1, its own custom language model. The move signals a strategic shift among applied AI companies away from exclusive dependence on frontier models toward proprietary inference infrastructure—a pattern driven by cost pressures and the fear that generic LLMs cannot adequately optimize for specialized workflows like visual app generation.

Training on Real-World Platform Data

According to TechCrunch AI, Base1 was developed and trained on a dataset sourced from tens of millions of authentic user interactions accumulated on Base44’s platform. This user-generated dataset represents a key competitive moat: as Base44 continues scaling, the training corpus expands automatically, creating a reinforcing loop that external model providers cannot replicate for this specific use case. Base44 CEO Maor Shlomo told the publication that owning the model “allows us a lot more optimizations on latency, cost, and efficiency”—a direct response to the rising inference costs that have become a material factor in unit economics for enterprise customers.

The Broader Defensibility Question

The decision reflects a wider conversation in venture capital about whether AI startups built exclusively on third-party models possess sustainable competitive advantages. According to Jonathan Userovici, a general partner at venture firm Headline (whose portfolio includes Mistral AI), defensibility for AI-native companies rests on three pillars: data, distribution, and tech stack. Base44 now controls two of the three directly; its distribution advantage stems from Wix’s installed base and the platform’s existing user loyalty.

However, Shlomo acknowledged that rivals with sufficient scale—such as Swedish unicorn Lovable—will eventually follow suit and train proprietary models. The more immediate threat may emerge from frontier labs themselves. Anthropic’s Claude Code, Cursor, and xAI’s Grok (both now owned by SpaceX) are entering the no-code and app-generation space, bringing with them the scale and data collection infrastructure to rapidly improve domain-specific performance. Userovici cautioned against overestimating the advantage of vertical specialization, citing Harvey, the legal-tech startup that abandoned its own model-training efforts in favor of relying on frontier models.

Why This Matters

Base44’s move signals a fork in the road for applied AI startups. If proprietary training on domain data proves to yield meaningfully better latency and cost profiles within 12 months, other well-capitalized platforms will follow, fragmenting the LLM market along vertical and use-case lines. Conversely, if frontier models continue to close the gap through architectural innovation and scale, startups like Base44 may find that the operational burden of maintaining an in-house LLM outweighs the efficiency gains—making continued dependence on external providers the rational choice, as Harvey’s pivot demonstrated. Enterprise customers making build-vs.-buy decisions for AI-augmented workflows should monitor whether Base1’s specialized training translates to measurable cost reductions in real deployments.