DBRX.
Open MoE model from Databricks Mosaic with native warehouse integration under MIT licence.
What it is.
Databricks's open-weights mixture-of-experts model, originally shipped by Mosaic AI in 2024 and extended through the Mosaic platform. MIT-licensed with 132B total parameters and 36B active per token. Trained on a Databricks-curated dataset of 12T tokens. Native integration with Unity Catalog, Delta tables, and the Mosaic AI Agent Framework.
Where it fits.
Databricks customers who want a frontier-class model fine-tuned on their warehouse without crossing a vendor boundary. The lakehouse-native fine-tune path is the differentiator — DBRX trains directly on Delta tables with Unity Catalog governance preserved. Cost-anchor in any procurement where the buyer already pays Databricks for compute.
- MIT licence removes commercial friction entirely
- Lakehouse-native fine-tuning on Unity Catalog data
- Mosaic AI platform handles serving and evaluation
- Behind Llama 4 and Qwen 3 on the latest open-source benchmarks
- Tightly coupled to Databricks for the full experience
Frequently asked.
What is the DBRX licence?
MIT. The most permissive licence in the open-weights tier. Commercial use is unrestricted with no MAU threshold.
Can DBRX be served outside Databricks?
Yes. MIT weights run on any serving stack. The differentiated path is Mosaic AI inside Databricks with warehouse integration intact.
How does DBRX compare to Llama 4?
Llama 4 leads on raw benchmark scores. DBRX wins for buyers already on Databricks who value lakehouse-native fine-tuning over the last percentage points of leaderboard performance.
