Particle.news

Base44 Rolls Out Base One, Its Specialized Model for Vibe Coding

Base44 says the fine-tuned model will cut inference costs, improve latency, tighten alignment for translating natural-language prompts into working web apps.

Overview

  • Base44 put Base One into production and began rolling it out to users, a development confirmed in coverage on June 29–30, 2026.
  • Base One is not built from scratch but is a fine-tune of an open-source foundation model trained with reinforcement-learning feedback from tens of millions of real platform interactions.
  • Base44 developed the model in collaboration with Wix’s machine learning and data science team and tuned it to the platform’s agentic harness of tools, instructions, and agent operations.
  • The company says the move is aimed at lowering inference costs and latency and producing outputs more closely aligned to vibe-coding tasks, which could make the platform faster and cheaper for customers.
  • Industry watchers see the launch as part of a wider trend where AI application companies weigh owning vertical models, data, and infrastructure for defensibility against relying on general frontier models from large providers.