DBRX Instruct is a Mixture-of-Experts (MoE) model with 132 billion total parameters, of which 36 billion are active during any single forward pass. Built by Databricks and released as an open model, DBRX was designed to match or exceed the quality of prior open-weight leaders (like LLaMA 2 and Mistral) while demonstrating that MoE architectures can be trained efficiently at scale. It is served via Together AI's inference infrastructure.
DBRX Instruct covers coding, mathematical reasoning, and general language tasks, with particular strength in software engineering contexts — reflecting Databricks' core user base of data engineers and ML practitioners. Its MoE structure provides a favorable quality-to-active-parameter ratio, meaning it reasons at 132B quality while only activating 36B parameters per token.
Key Features
132B total parameters with 36B active via Mixture-of-Experts routing
Strong code generation and software engineering task performance
Mathematical reasoning competitive with larger dense models
Open-weight model allowing self-hosting and fine-tuning
Efficient inference due to sparse MoE activation pattern
Multi-turn instruction following for complex analytical tasks
Ideal Use Cases
Code generation and debugging for data engineering and ML workflows
Self-hosted enterprise deployments where model weights must stay internal
Research into open MoE architectures and sparse activation models
Complex question answering and document analysis pipelines
Mathematical and scientific reasoning in academic or enterprise contexts
Example Prompts for DBRX
Technical Specifications
| Provider | Together |
| Category | Text |
| Modality | Text -> Text |
| Context Window | 32K tokens |
Frequently Asked Questions
Try DBRX now
Start using DBRX instantly — 100 free credits, no credit card required. Access 750+ AI models through one platform.