Skip to main content

CanopyWave Flower Animation Roundup: 4 Models Compared

CanopyWave Flower Animation Roundup: 4 Models Compared

We asked four models on CanopyWave for a looping CSS flower animation. Half of them actually looped.

bots.so

One prompt. Four models. All on CanopyWave. The prompt asked for a flower growing from brown dirt as an animated loop using only CSS. Simple brief — but half the models ignored the “loop” part.


Prompt

“Create a gorgeous flower growing from brown dirt, as an animated loop using only HTML and CSS.”

Pure HTML/CSS · No JavaScript · Single file

DeepSeek V3.2 CanopyWave
Tokens
1,991
Lines of Code
245
MiniMax M2.5 CanopyWave
Tokens
4,161
Lines of Code
368

Model A

DeepSeek V3.2

CanopyWave

vs

Model B

MiniMax M2.5

CanopyWave

1,991
Tokens
4,161
245
Lines of Code
368

Green highlights the leading metric per row. Fewer tokens = more efficient. More lines = more detail.

Additional models

Kimi K2.5 CanopyWave
Tokens
3,268
Lines of Code
208
MiMo V2 Flash CanopyWave
Tokens
2,119
Lines of Code
234

The loop test

This is the single most important question: did the model build an animation that loops?

ModelLoops?How
Kimi K2.5Yes8-second infinite cycle, percentage-based keyframes, grow-bloom-fade-reset
MiniMax M2.5Yes8-second infinite cycle, chain-of-thought planned the loop architecture
DeepSeek V3.2NoGrowth uses forwards (one-shot), only sparkles loop
MiMo V2 FlashNoGrowth uses forwards (one-shot), only petal wiggle loops

Two out of four followed the prompt. The other two built a grow-then-idle sequence — the flower appears once and stays there. That’s not a loop.

Prompt adherence scorecard

RequirementDeepSeek V3.2MiniMax M2.5Kimi K2.5MiMo V2 Flash
Flower with petals8 petals8 petals8 petals4 petals
Brown dirtYesYesYes (textured)Yes (bouncing)
Grows from dirtYesYesYesYes
Animated loopNoYesYesNo
CSS only (no JS)YesYesYesYes

Every model nails the visual basics — flower, dirt, growth, CSS-only. The differentiator is whether they read “animated loop” as a requirement or a suggestion.

Bottom line

Kimi K2.5 and MiniMax M2.5 followed the prompt. DeepSeek V3.2 and MiMo V2 Flash built nice flowers but didn’t build loops. If you’re evaluating models for instruction-following on creative tasks, this is the kind of detail that matters.


Part of our Model Animation Benchmark series on CanopyWave. Subscribe for new comparisons.

Newsletter

Get the signal, skip the noise.

Weekly digest of new models and provider updates across 40+ compute providers. Curated for AI builders who ship.

New model releases
Capability updates
Provider status
bots.so
The AI Inference Model Index
© bots.so — The AI Inference Model Index

bots.so aggregates publicly available model deployment information from official provider sources. We are not affiliated with any model provider. Model availability changes rapidly; always verify on official sites.