This is one of those features that exists before most people realize it exists.
A very new capability in Microsoft Fabric now allows Direct Lake semantic models to run in mixed (composite) mode — and that changes the game.
I’ve just published a video where I show how to do it step by step.
But here’s why this matters 👇
⚡ Why combining Direct Lake + Composite Models is powerful
🔹 Aggregations finally make sense with Direct Lake
Pre-aggregated Import tables can accelerate high-level queries while keeping detail in Direct Lake
🔹 Performance where it matters most
Hot paths hit in-memory aggregations, cold paths stay in the lake — best of both worlds
🔹 More flexible modeling patterns
You’re no longer forced into a single storage mode decision for the entire model
🔹 Scalability without duplicating everything
Keep large fact tables in OneLake while selectively importing only what adds value
🔹 A smoother transition path
Ideal for evolving existing Import or Direct Lake models without redesigning everything
The video focuses purely on how to implement this, but the implications go far beyond the demo.
watch?v=gwmXwtVYwy8
