DBRX MoE-B
LLMLLM
Mid-size DBRX family member: 23.5B total parameters, 6.6B active. Used to study MoE training efficiency. Achieves 45.5% on the Databricks Gauntlet with 1.7x fewer FLOPs than LLaMA2-13B (13B active parameters).
Technical specification
Context window
Parameters
License
Tools
Fine-tuning
Weights access
Hardware requirementsInternal Databricks research model; no public checkpoint available.
Last updated: May 4, 2026
Modalities
Input
Text
Output
Text
Code