Databricks: Data Platforms, MLflow and DBRX Models

A platform connecting data and models

Databricks dates its founding to 2013 and identifies founders associated with Apache Spark, Delta Lake, MLflow and Unity Catalog. Its data and AI platform joins processing, governance and model work. That retrospective account of the founders’ projects does not mean every named project existed when the company began. [1]

A substantive model contribution

The March 2024 DBRX release included base and instruction-tuned model weights, plus API access for Databricks customers. Its research team described a mixture-of-experts architecture with 132 billion total and 36 billion active parameters. Vendor benchmark comparisons depend on their test configurations. [2]

The Databricks Open Model License governs DBRX and contains use restrictions, including on improving other LLMs and a separate permission requirement above 700 million monthly users. The weights release should therefore be described with its actual custom terms, rather than generalized as Apache or MIT licensed. [3]