Base44 is training its own LLM — and thinks it'll beat GPT-4
vibe coding platform base44 launches its own llm, base1, trained on tens of millions of user interactions to cut latency and cost. it signals a shift where ai startups build proprietary models to defend against frontier labs.
base44 is launching base1, a proprietary large language model developed using a dataset of tens of millions of platform interactions, aiming to surpass frontier models like opus in specific coding tasks. the company cites latency, cost efficiency, and output optimization as primary drivers, responding to enterprise demands for roi where general models often overspend on resources. this move highlights a strategic pivot where applied ai companies leverage proprietary data for defensibility against encroaching tools like cursor and claude code. while vc investor jonathan userovici notes some firms like harvey abandoned custom training, base44 bets specialized models will outperform generalized ones. the rollout suggests a bifurcating market where vertical ai stacks compete on infrastructure ownership.