在 Databricks SQL 中运行开源决策模型 SemIf-OpenJev

• Run models like SemIf-OpenJev directly on governed Databricks data using serverless GPUs and managed Model Serving.
• Query custom model endpoints from SQL or Lakeflow jobs with ai_query, returning structured decisions and probability scores at scale.
• Get started with an importable notebook that deploys the model in a few steps, then adapt it to your own data, decision criteria, and use cases.

Over the weekend, “System One” decision models such as Jev have launched, which are a growing class of foundation models that are able to produce well-calibrated decisions from a discrete set of options. These models are usually extremely fast and also cheap, allowing a broad range of applications on large data. Potential use cases range from analyzing customer support transcripts, to applying complex decisions at scale across your data.

The open source community has also been busy launching open weight versions of these decision models, from Smelf-open-jev, to Laya, to Kev. But the power of these models are only unlocked when they can touch large amounts of data. Today we are excited to share how you can serve open weight decision models on Databricks, and run them directly on your governed data. You can even access these models directly from your SQL console or production jobs in Lakeflow via ai_query. Moreover, products like AI Runtime enable you to customize or post-train these models to tailor them to your specific enterprise context.

In this post, we’ll show how you can use such a model, SemIf-OpenJev, directly in Databricks, to classify hotel reviews as good or bad.

From notebook to endpoint

We’ve packaged the workflow into an importable Databricks Notebook.

Getting started takes three steps:

1. Pick the model name/schema
2. Select Serverless GPU
3. Click Run All

The notebook uses Databricks AI Runtime to provide serverless, on-demand GPU compute without requiring you to set up or manage GPU infrastructure.

From there, it:

• Downloads the SemIf-OpenJev model files.
• Logs and registers the model using Express Deployments.
• Creates a GPU Model Serving endpoint automatically.

At that point, you have the open source classifier running behind a managed endpoint that you can call from applications — or directly from SQL.

Call the model from SQL with ai_query

One useful part of this setup is that the endpoint doesn’t need to expose a standard chat-completions interface.

Databricks ai_query supports arbitrary custom model APIs, so you can send SemIf’s structured request format directly from SQL.

For example:

Now you can apply the model to reviews already stored in Databricks and get the selected classification and option probabilities back as structured columns.

Decision models are moving quickly, and open source implementations like SemIf-OpenJev make it much easier to experiment with the pattern yourself.

Import the notebookdirectly into your Databricks workspace, select Serverless GPU, and click Run All to get started. From there, try it on your own data, swap in your own decision criteria, and see where the pattern fits in your applications.