DBRX

DBRX was Databricks' statement release in March 2024: a 132B-parameter mixture-of-experts LLM with 36B active parameters on any input, pretrained on 12T tokens of text and code with a 32K maximum context, using a fine-grained MoE architecture built on the MegaBlocks research project. Base and Instruct variants are downloadable from Hugging Face under the Databricks Open Model License and an accompanying acceptable use policy, which allow commercial use subject to their terms, and the model is also distributed through the Databricks Marketplace. DBRX Instruct specializes in few-turn interactions and was aimed squarely at enterprise workloads, with Databricks publishing benchmarks against contemporary open models at launch. Running it is a serious infrastructure exercise given the parameter count, and the original GitHub repository has since been taken down, leaving the Hugging Face model cards as the primary artifact. The release matters historically as one of the first demonstrations that an enterprise data platform vendor could train a frontier-scale open-weight MoE model entirely in-house.