Databricks · General LLM

DBRX

Databricks' open-source mixture-of-experts language model designed for enterprise data intelligence and analytics workflows.

Overview

DBRX is Databricks' open-source large language model built on a fine-grained mixture-of-experts architecture with 132B total parameters but only 36B active per token. It was the most capable open-source model at the time of its release, outperforming Llama 2 70B and Mixtral on standard benchmarks. DBRX is tightly integrated with the Databricks Lakehouse platform, making it a natural choice for organizations already using Databricks for their data and analytics infrastructure.

Parameters

132B total, 36B active (MoE)

Context Window

32K tokens

Architecture

Fine-grained MoE (16 experts, 4 active)

Training Data

12T tokens

License

Databricks Open Model License

Capabilities

General-purpose text generation and reasoning

Code generation and SQL query writing

Data analysis and business intelligence

Fine-grained mixture-of-experts efficiency

Use Cases

Building AI-powered data analytics within Databricks Lakehouse

Generating SQL queries and data transformations from natural language

Creating enterprise chatbots grounded in organizational data

Automating data documentation and metadata management

Pros

  • +Strong open-source model with transparent training details
  • +Efficient MoE architecture reduces inference costs
  • +Deep integration with Databricks data platform
  • +Competitive performance for its active parameter count

Cons

  • -Large total parameter count requires significant memory
  • -Custom license is more restrictive than Apache 2.0 or MIT
  • -Best experienced within the Databricks ecosystem
  • -Surpassed by newer open models like Llama 3

Pricing

Free to download. Available through Databricks Model Serving with pay-per-token pricing. Self-hosting requires multi-GPU infrastructure.

Related Models