Classification-Only AI Models Challenge LLMs for Bulk Data Labeling in Lakehouses

Teams using large language models for text classification in data pipelines face high costs and slow performance, particularly when processing tens of millions of rows in Apache Iceberg lakehouses. In September 2026, a company called TypeSafe AI released Jev, a model designed exclusively for classification tasks such as labeling, scoring, or returning yes/no probabilities with calibrated confidence. Unlike LLMs, which generate output token by token and often return fragile JSON, dedicated classification models complete tasks in a single forward pass, cutting both latency and expense. Several open-source projects quickly emerged to replicate Jev's interface on self-hostable models. The article explores how these two model types can be used together within an Iceberg lakehouse, using Dremio's SQL AI functions as one practical example.