DBRX Base

DBRX Base is a state-of-the-art MoE language model designed by Databricks to maximize quality and efficiency at scale. A router selects a small subset of experts per token (top-k routing), delivering high throughput with frontier-level accuracy while controlling latency and cost. The model supports long inputs (typically around a 32K context window) and is trained on a carefully filtered blend of code, math, technical web, and high-quality prose so it generalizes well to analysis and software tasks.

As a base model, it’s unaligned by default—meant to be adapted. Teams commonly:

Fine-tune with LoRA/DPO/RL to add style, tools, or policies

Ground with RAG for factual work over private corpora

Constrain outputs to JSON/DSLs for agents and automations

DBRX integrates smoothly with common stacks (vLLM/Transformers) and Databricks Mosaic AI for training/serving, with quantization options for cheaper inference. Use DBRX Base when you want maximum control over behavior; choose DBRX Instruct when you need a ready-to-use conversational model.