Semantic model

The tables, relationships and measures that turn raw data into the business numbers every report uses.

The semantic model is the part of Power BI most people never see, and the part that decides almost everything: whether reports are fast, whether the numbers match, and whether Copilot gives sensible answers.

Part of the Model layer of the ecosystem. Reviewed 13 November 2026 by Piyush Jain.

What Semantic model is

A semantic model (it used to be called a dataset) holds your tables, the relationships between them and the measures that calculate the business numbers. Reports do not calculate anything on their own; every visual asks the model.

The model also decides how data is stored. In Import mode it is compressed into memory; in DirectQuery each visual queries the source; in Direct Lake it reads Delta tables in OneLake. Security rules, such as row-level security, live here too.

When to use it

Use it when

  • Always: every Power BI report has one. The real decision is how many you have.
  • Share one model across many reports when they report on the same business process, so each number has one definition.

Think twice when

  • Avoid building a new model for every report. That is how the same measure ends up with three definitions.
  • Avoid one giant model for the whole company. Split by business process when models get hard to own.

The mistakes I see most

From opening other people's models. Each one is fixable.

  1. 1

    One flat, wide table

    A single table with every column is easy to start and painful forever: big, slow and hard to filter correctly. A star schema, with fact tables for events and dimension tables for who, what and when, is smaller, faster and gives correct totals.

  2. 2

    Two-way relationships everywhere

    Bi-directional filters feel like they fix a problem in one visual, then cause wrong numbers or slow queries elsewhere. Keep relationships single-direction by default and handle the exceptions in DAX.

  3. 3

    Columns nobody uses

    IDs, GUIDs and timestamps to the second, imported just in case, can make up most of a model's memory. Removing one unused high-cardinality column often shrinks a model by a third.

Questions people ask

Is a semantic model the same as a Power BI dataset?
Yes. Microsoft renamed datasets to semantic models in 2023. The thing itself did not change.
Should every report have its own semantic model?
No. Reports on the same business process should share one model, so every measure has one definition. Separate models make sense when the data, the owners or the security rules are different.
Why does the semantic model matter for Copilot?
Copilot reads the model's table, column and measure names, descriptions and relationships to work out what your data means. A clear, well-named model gets far better answers.

Sources: Microsoft Learn: Star schema guidance for Power BI. Opinions are my own.