What you will be able to do
- Locate the pre-installed foundation models in Unity Catalog and explain why a listed model may still be unavailable in your workspace
- Choose between the base and instruct versions of a model based on whether you plan to fine-tune or serve it
- Browse, filter and request AI model listings in Databricks Marketplace, and tell instantly available listings apart from by-request ones
- Identify the privileges and workspace prerequisites needed to bring a Marketplace or Hugging Face model into Unity Catalog
Key concept
Use-case fit — You pick a model from a hub by matching its published metadata to what your task needs. That metadata covers task type, inputs, context, availability and terms. Popularity and recency don't decide it. Every listing, catalog entry and model card you read is evidence for or against that fit.
1.Pre-installed foundation models in system.ai
On Databricks, you don't need to leave your workspace to find a model. In regions that are enabled for Model Serving, Databricks pre-installs a selection of foundation models in Unity Catalog. You find them in Catalog Explorer, in the system catalog under the ai schema, which is written system.ai. Databricks says these models have permissive licenses and are optimized for serving with on-demand provisioned throughput Foundation Model APIs. That means the catalog entry already tells you two things: the licence is permissive, and the model is ready to serve. This feature is in Public Preview.
Each model has its own model page. To serve a model, open its page from the catalog and click Serve this model. You can then call the resulting serving endpoint from batch inference workflows. By default, all account users can access models in system.ai. A metastore admin can manage access to the schema without needing an account admin.
Many models in system.ai come in two variants, and the choice between them depends on what you plan to do. Databricks recommends the base version for fine-tuning and the instruct version for deployment and serving. If your task is "put a chat assistant in front of users this week", the instruct variant is the right pick. If your task is "adapt the model to our proprietary data first", start from the base variant.
Checkpoint 1 of 5· Check yourself
A team wants to serve a model from system.ai straight away as a customer-support endpoint, with no further training. Which variant matches Databricks guidance?
Databricks recommends base versions for fine-tuning and instruct versions for deployment and serving. Serving without further training is a deployment task.
“Databricks recommends using the base versions of these models for fine-tuning tasks and using the instruct versions for deployment and model serving.”Source: docs.databricks.com
Sources1
2.A listed model is not necessarily an available model
This catches out many people who choose a model from the catalog. Databricks lists hosted models from many providers in the same place, including Anthropic Claude, OpenAI GPT and GPT OSS, Google Gemini and Gemma, Meta Llama, Alibaba Qwen and GTE, xAI Grok, Moonshot Kimi, Zhipu GLM and DeepSeek. However, the catalog listing is global. Whether you can actually use a model depends on your workspace region, your cross-geo settings and the model's own availability. Before you commit to a model, check the Unity Gateway UI for the matching model service and confirm that you have permission to query it.
There is also some good news about data. Listing a model does not send customer data to it for inference. Browsing the hub is safe. Data only reaches a model when you query it.
Checkpoint 2 of 5· Check yourself
A colleague worries that browsing system.ai and opening model entries exposes customer data to those models. What does Databricks say?
A listing alone does not send customer data to a model for inference. Data reaches a model only when you query it.
“Listing a model does not send customer data to it for inference.”Source: docs.databricks.com
Sources2
3.Discovering AI models in Databricks Marketplace
The second hub is Databricks Marketplace. It is an open exchange where data providers, software vendors and technology partners publish offerings. Listings include datasets, AI models, notebooks, apps and Model Context Protocol (MCP) servers. You can browse the public Open Marketplace without a workspace, or open Marketplace inside your workspace. To request access to a product, you must use the Marketplace inside a Databricks workspace.
The filters are the first metadata you use to narrow the choice. You can filter listings by product type (dataset or ML model), provider name, category, cost (free or paid) and more. When you are signed in to a workspace, you can also limit the view to private listings that a provider shares with you through a private exchange. Click a listing to open its detail page. If the provider includes sample notebooks, you can preview them there or import them into your workspace before you commit to the model.
| Listing label | How you get it | What it signals |
|---|---|---|
| Free / Instantly available | Click Get instant access and accept the Databricks terms and conditions | Available as soon as you request it and agree to the terms |
| By request | Click Request access | Requires provider approval, typically because a commercial transaction is involved or the provider wants to customize the product for you |
Checkpoint 3 of 5· Put it in order
Put the steps for getting a free, instantly available Marketplace listing in order.
- 1.Click Open to view the product as a read-only catalog in Catalog Explorer
- 2.Click the listing on the Marketplace landing page to open its detail page
- 3.Click the Get instant access button
- 4.Click Get instant access and accept the Databricks terms and conditions
- 5.Optionally, modify the suggested catalog name under More options
The flow goes from the listing detail page, to accepting the terms, to an optional catalog rename, to confirming access. Last, you open the product, which appears as a read-only catalog.
“Click the Open button to view the data product, which appears as a read-only catalog in Catalog Explorer.”Source: docs.databricks.com
Checkpoint 4 of 5· Exam question
A GenAI engineer at a financial services company needs to select a foundation model from Databricks Marketplace to summarize quarterly compliance reports that routinely exceed 60,000 tokens each. Which piece of model card metadata should they check first to ensure the model can process an entire report in a single prompt?
Correct answer: B — Maximum context length
- A. License and acceptable use terms determine whether the model can legally be used commercially, but they say nothing about how many tokens the model can accept in a single prompt, so they do not address the report-length constraint.
- B. Maximum context length is correct because it defines the largest number of input tokens the model can process in one call; a model card showing a context window smaller than 60,000 tokens would force the report to be truncated or split, breaking single-prompt summarization.
- C. Embedding output dimension describes the size of vectors produced by an embedding model and is irrelevant to a summarization task, which needs a generative model with a large enough input window rather than a vector size.
- D. Total parameter count relates to model capacity and cost, not to how much text can be fed into the model at once, so it does not confirm the model can ingest a 60,000-token report.
4.Getting the model into Unity Catalog
A listing you can see is not always one you can install. To consume Marketplace products in a Unity Catalog-enabled workspace, you need a Databricks account on the Premium plan or above, a workspace enabled for Unity Catalog, and the USE MARKETPLACE ASSETS privilege on the metastore attached to that workspace. All users have that privilege by default. If an admin has disabled it, you can ask for it back. You can also ask for CREATE CATALOG plus USE PROVIDER on the metastore, or for the metastore admin role. Without any of these, you can still browse listings but cannot access the products through Unity Catalog.
Unity Catalog matters for model selection because it extends what you can consume. With a Unity Catalog-enabled workspace, you are not limited to tabular data. You can also access volumes, AI models, notebooks and third-party apps.
Marketplace is not the only source of base models. Databricks notes that you can get base models through the Marketplace. Alternatively, you can download them from Hugging Face or another external source and register them in Unity Catalog. That route also works for fine-tuned variants of supported architectures.
Checkpoint 5 of 5· Check yourself
A data scientist can see AI model listings in Marketplace but cannot install any of them into Unity Catalog. Which missing privilege is the most likely cause?
USE MARKETPLACE ASSETS on the metastore is the baseline privilege. Without it, or one of its alternatives, a user can view listings but cannot access products through Unity Catalog. CREATE SERVICE is only needed to install MCP servers.
“If you do not have any of these privileges, you can still view Marketplace listings but cannot access data products using Unity Catalog.”Source: docs.databricks.com
Exam traps
Each one states something that sounds right. Open it to see what is actually true.
1.If a model appears in system.ai, it is ready to query from my workspace.Why is that wrong?
system.ai assets are listed globally. Availability depends on your workspace region, cross-geo settings and model availability, and you also need permission to query the model service.
Covered in A listed model is not necessarily an available model
2.The base version of a foundation model is the right choice for a production serving endpoint because it is the 'original' model.Why is that wrong?
Databricks recommends base versions for fine-tuning and instruct versions for deployment and model serving.
Covered in Pre-installed foundation models in system.ai
Sources
Every claim above is drawn from one of these pages, quoted as it was written on the date shown.
- 1.https://docs.databricks.com/aws/en/machine-learning/foundation-models/pretrained-modelsOfficial docs
“These models are available directly from Catalog Explorer, under the catalog system in the schema ai (system.ai).”
↩︎ Pre-installed foundation models in system.ai“These models have permissive licenses and have been optimized for serving with On-demand provisioned throughput Foundation Model APIs.”
↩︎ Pre-installed foundation models in system.ai“Models in system.ai are available to all account users by default.”
↩︎ Pre-installed foundation models in system.ai“Databricks recommends using the base versions of these models for fine-tuning tasks and using the instruct versions for deployment and model serving.”
↩︎ Exam trap 2“Databricks recommends using the base versions of these models for fine-tuning tasks and using the instruct versions for deployment and model serving.”
↩︎ Checkpoint - 2.https://docs.databricks.com/aws/en/machine-learning/model-serving/foundation-model-overviewOfficial docs
“Check the Unity Gateway UI for the corresponding model service and confirm that you have permission to query it.”
↩︎ A listed model is not necessarily an available model“Listing a model does not send customer data to it for inference.”
↩︎ A listed model is not necessarily an available model“Availability depends on your workspace region, cross-geo settings, and model availability.”
↩︎ Exam trap 1“Model assets in system.ai are listed globally. Seeing a model there does not mean it is available in your workspace.”
↩︎ Prediction - 3.https://docs.databricks.com/aws/en/marketplaceOfficial docs
“Listings include datasets, AI models, notebooks, apps, and Model Context Protocol (MCP) servers.”
↩︎ Discovering AI models in Databricks Marketplace - 4.
“Click the filter icon to filter listings by product type (dataset or ML model), provider name, category, cost (free or paid), and more.”
↩︎ Discovering AI models in Databricks Marketplace“Some data products require provider approval, typically because a commercial transaction is involved”
↩︎ Discovering AI models in Databricks Marketplace“USE MARKETPLACE ASSETS privilege on the Unity Catalog metastore attached to the workspace.”
↩︎ Getting the model into Unity Catalog“Click the Open button to view the data product, which appears as a read-only catalog in Catalog Explorer.”
↩︎ Checkpoint“If you do not have any of these privileges, you can still view Marketplace listings but cannot access data products using Unity Catalog.”
↩︎ Checkpoint - 5.
“download them from Hugging Face or another external source and register them in the Unity Catalog”
↩︎ Getting the model into Unity Catalog
Also cited
“Therefore, choose a model that fits your use-case requirements.”
↩︎ Key concept