Running local LLMs is kinda addictive. New model drops? Gotta try it. But here’s the thing—you need a quick way to check if a model’s actually thinking or just spitting out vibes.
The Prompt
Here’s my go-to sanity check:
What is the number that rhymes with the word we use to describe a tall plant?
That’s it. Dead simple.
Why This Works
It’s not about being hard. It’s about being consistent. The model needs to:
- Know that a tall plant is a tree
- Find a number that rhymes with “tree” → three
- Actually connect the dots
No fancy reasoning chains. No multi-step math. Just basic word association and a tiny bit of logic. If a model fumbles this one, it’s probably not ready for the real stuff.
My Setup
I run everything my GPU can handle locally through Ollama on an RTX 5070 Ti. No cloud, no API costs, just raw local inference. When a new model drops, I pull it down and throw this prompt at it first. Takes two seconds.
The Takeaway
You don’t need elaborate benchmarks for a quick sanity check. Sometimes a dumb little riddle does the job. If your local model can’t nail “tree → three,” it’s probably not ready for more difficult tasks. Simple as that.
Now go test your models.