Why we made MiniMax M3 the default model

The default model behind Ullbek changed again. Four reasons, and one moment that decided it — the hero section.

The default model behind Ullbek has changed again. A while back, it was GLM 5.2. Today, it's MiniMax M3. That's not a brand change — that's how a moving target works. New models arrive, we re-evaluate, and the best one takes the seat.

This is the short version. Four reasons it took the seat, and one moment that decided it.

Working with a new model is a bit like interviewing a candidate. You hand it small tests — a brief, a critique, a constraint — and you watch how it reasons. Most pass. A few impress. Occasionally one surprises you on something ordinary. M3 surprised us on a hero section.

Four reasons it took the seat

01 · Vision

It can see

M3 is vision-native. It doesn't only read your prompt — it looks at the images already on the page, understands what's there, and makes decisions with that information. When you say the hero feels cramped, it can see why. Most models stop at the text.

02 · Taste

It gets the brief under the brief

Ask for a "minimalist portfolio" and most models pick a font and a margin. M3 decides which kind of minimalist — gallery, editorial, technical, organic — and commits to it. The brief you wrote was a starting point, not a constraint. The model fills in the parts you didn't say.

03 · Performance

Taste at price

It produces output that looks like a small studio was paid for it, at a fraction of the inference cost. The price-to-performance is a different league from what we had. That matters because it means every site you build can have that level of craft — not just the ones where you spend extra.

04 · Density

The right balance

Not too dense that it feels crowded. Not so sparse it feels lazy. M3 lands in the band where every section earns its place. The hardest thing to teach a model is restraint, and M3 has it — every paragraph, every card, every gap on the page is intentional.

The moment we knew

The aha moment

"The image was cropped perfectly. Even the aspect ratio was preserved."

Step 01
We asked the model to restructure a hero section on a real site we were building — a small photography portfolio. The original layout was fine, just a little cramped against the new header.
Step 02
It understood it wasn't a tweak — it was a different layout. The copy needed to move, the image needed to breathe, the proportion had to change. So it changed the proportion first.
Step 03
It cropped the existing image to fit. The crop kept the subject, the framing was right, and the aspect ratio was preserved. Nothing was stretched, nothing was duplicated, nothing was invented from scratch.

Most models would have asked which image to use, or stretched the original, or broken the framing. M3 looked at the image, decided what mattered, and acted on it. That's when we knew this model was different.

When we looked at the result, the page didn't feel "regenerated." It felt rearranged — by someone who'd actually seen it. The new hero sat next to the new header like they'd always been planned together. That's the part that got us.

What it changes for the sites you build

Switching the default isn't a marketing move — it's a quiet one that ripples through every site you make. Three small differences we kept seeing during testing, with no extra prompting required:

Before

When a prompt mentions an image, most models wait for permission — or stretch the original to fit the new frame.

M3

It looks at the image, decides what matters, and crops it. The subject stays, the framing stays, the aspect ratio stays.

Before

When a brief is vague, most models hedge — three fonts, two palettes, a layout that doesn't quite decide what it wants to be.

M3

It picks a direction and commits. The site comes out with a point of view, not a menu of options.

Before

When you ask for a "restructure," most models do the minimum edit you asked for — they don't reason about the intent underneath.

M3

It reads past the words. A restructure is treated as a restructure — proportion, balance, the whole layout shifts, not just one knob.

None of these are features you have to learn. They're just the way the model behaves in the background, on every prompt. The result is that sites come out closer to what you meant the first time, with fewer follow-ups.

The default is a moving target

On not anchoring to a model

GLM 5.2 was the default for a while, and the same evaluation method we ran for it is what we ran for M3 — the prompt, the live site it produced, the open scorecard. When the next model arrives, and one always does, we'll run the same evaluation, and the seat might move again.

The default will keep changing. The method doesn't. You don't anchor to a model, you anchor to a method. That's how a moving target stays useful instead of just moving.

Read the GLM 5.2 evaluation →  ·  How we evaluate models →

See what M3 builds for you

Type a sentence. Hit enter. The first $2 of credits is on us.