Choosing a model from a hub or marketplace using its metadata or model card

When choosing a foundation or embedding model for a use case, use Databricks Marketplace and the Unity Catalog-integrated model listings to compare model cards before committing to one. Model cards summarize the license, intended use, context length, input/output modalities, and benchmark performance, which should be matched against the task's latency, cost, and quality requirements. Selecting the right model up front avoids costly rework later when serving, fine-tuning, or governance constraints surface.

1 · Learn the must-know

• 1Model cards on Databricks Marketplace and in Unity Catalog list license terms, intended use cases, and known limitations, so check licensing compatibility (e.g., commercial vs. non-commercial) before deployment.
• 2Context window size and max token limits in the model card determine whether a model can support the RAG or prompt design pattern the application needs.
• 3Benchmark metrics (e.g., accuracy, latency, throughput) provided in the card should be compared against task requirements rather than assuming the largest model is best.
• 4Smaller or task-specific models (e.g., an embedding model vs. a general chat/instruct model) are often preferable for narrow tasks due to lower cost and latency.
• 5Foundation Model APIs expose curated, Databricks-hosted models with pay-per-token or provisioned throughput options, and their model cards indicate which serving mode is supported.
• 6Models registered via Unity Catalog carry lineage and governance metadata, so selecting a Unity Catalog-registered model simplifies permissions, tracking, and reproducibility compared to ungoverned external sources.

2 · Check your understanding