Blender 4.5 added a third boolean solver, called Manifold. It is sold as the fast one, and it is faster than Exact.
It is also the one that refuses to run on most of the meshes I cut with. So I cut twenty-six of my own objects with it and counted.
Twenty-two of those tests gave a clear answer. Thirteen of the twenty-two were refused.
What the solver will accept
The manual is short about this. Manifold works only on manifold meshes, plus the special case of subtracting a plane.
Manifold means every edge has exactly two faces. No open shells, no loose vertices, no edge shared by three faces.
The same page says the other solvers usually cope with messy input. This one does not cope at all.
The test
I took three sources. A hard-surface pistol built as fourteen parts, a stylised barrel of eight objects, and four meshes from a door knocker set I sell.
Each object got the same treatment. One cube at the centre of its bounding box, Difference, in Blender 5.1.1.
I ran Exact first, then Manifold. If Exact cut it and Manifold left it untouched, I counted that as a refusal.
Thirteen of twenty-two
Four tests told me nothing, because the cube missed and Exact did nothing either. That leaves twenty-two real cuts.
Manifold refused thirteen of them. The pistol was the worst file by a long way.
Ten of its fourteen parts were refused. Only two parts came out cut.
That is not sloppy modelling. Panels are open shells on purpose, because nobody renders the back of a slide plate.
The knockers all passed
All four door knocker meshes worked. They come from a sculpt and a retopology pass, so they are closed by construction.
You can see the split in one line. Sculpting gives you watertight meshes, and box modelling gives you the meshes booleans are for.
That is the awkward part. The solver is best at the geometry least likely to need it.
Manifold edges are not enough
Here is the result I did not expect. Four of my thirteen refusals had no non-manifold edges at all.
The pistol has one small part with 586 faces and 1,370 edges, every edge shared by two faces. Exact cut it and gave back 595 faces. Manifold gave back the same 586, with no error message at all.
So you cannot pre-check with the edge count alone. Passing that test does not mean the solver will run.
The speed, on one real mesh
I used one door knocker at 46,017 faces, cut by a closed cube. A 24-core i9-14900HX, median of three runs.
In Blender 4.5.3 LTS:
- Float: 0.045 seconds
- Manifold: 0.072 seconds
- Exact: 0.221 seconds
In Blender 5.1.1, same file:
- Float: 0.040 seconds
- Manifold: 0.065 seconds
- Exact: 0.193 seconds
Manifold is about three times faster than Exact. It is also about sixty per cent slower than Float.
What failure looks like on screen
I deleted one face from the cutter cube and ran it again. Manifold returned the mesh unchanged, in 0.027 seconds.
Float and Exact cut it anyway. Only Manifold stopped, saying it cannot execute with non-manifold inputs.
In the viewport this reads as nothing happening. You drag the cutter and the surface never opens.
Two smaller things
Exact did get faster between versions, from 0.221 to 0.193 seconds. That is about twelve per cent, not the thirty-five the notes quote.
The bigger figure is for meshes carrying many attributes. Mine is a game mesh with one UV map.
The Python name of the Float solver also changed in 5.0. Add-ons written for 4.5 set a value that no longer exists.
Normals are a separate problem
Flipped normals break booleans in every solver. Report 160436 was filed in June 2026 and is still open.
So a mesh can be closed and still cut wrong. Recalculate outside before you blame the solver.
What I do now
- Try Manifold on the part in front of you. It either cuts or it does nothing, and you know in one second.
- Keep cutters closed. A plane with Solidify is fine, an open box is not.
- Leave Exact as the default for everything else.
That last line is the honest ending. Exact is slower and it finishes the job on meshes Manifold will not open.
It also changed what I check before I ship. A base mesh that is not watertight is a base mesh nobody can cut with.
So the count is going onto my own release list, and onto the add-ons I maintain. Run it on one of your own prop sets first.
