Nano Language Models

Can a 1.7B AI Model Really Write Excel Formulas? Let’s Test It.

Most people assume that only the biggest models — GPT-4, Claude 3.5, Gemini — can handle structured reasoning tasks like Excel formulas.

So here’s a question:

What happens when you give the job to a tiny 1.7B-parameter model instead?

That’s what 6SigmaMind is built to explore.

Small models shouldn’t be good at Excel.
But… sometimes they are.
And when they are, it feels like discovering a cheat code.

Today, you get to try it yourself.

👉 Test 6SigmaMind live: https://huggingface.co/spaces/benkemp/6SigmaMindv2

🔍 Why This Experiment Matters

The AI world is shifting.
Big models are powerful — but small models are:

  • faster
  • cheaper
  • easier to embed
  • easier to run privately
  • easier to fine-tune
  • more fun to test because they occasionally surprise you

6SigmaMind is meant to answer one question:

How far can a single-purpose small model go if you aim it at one task — Excel logic?

The result is surprisingly capable for its size.

Let’s test that.

🧪 Test #1: SUMIFS and COUNTIFS

Try these prompts in the demo:

  • “Sum all values in column C where column B equals ‘Closed’.”
  • “Count how many cells in A2:A200 are greater than 50.”
  • “Return the number of empty cells in column F.”

Typical answers look like:

=SUMIFS(C:C, B:B, "Closed")
=COUNTIFS(A2:A200, ">50")
=COUNTBLANK(F:F)

A tiny model producing structured formulas?
That’s already a win.

🧪 Test #2: XLOOKUP + INDEX/MATCH

Ask:

  • “Give an Excel formula that returns the price in D where the SKU in A matches H2.”
  • “Use INDEX/MATCH to return a value from column C where A equals F2.”

Expected output:

=XLOOKUP(H2, A:A, D:D)

or:

=INDEX(C:C, MATCH(F2, A:A, 0))

This is the area where small models start showing personality —
sometimes perfect, sometimes creative, sometimes… interesting.
That’s the point of the experiment.

🧪 Test #3: Logical formulas (IF, AND, OR)

Give it these:

  • “If A2 > 100 return ‘High’ else return ‘OK’.”
  • “Check if B2 is blank.”

Expected outputs:

=IF(A2>100, "High", "OK")
=IF(B2="", TRUE, FALSE)

It often nails these on the first attempt.

🧪 Test #4: Statistics (the fun part)

Small models struggle with statistics — so this is where you can break it.

Try:

  • “Calculate the standard deviation for B2:B80.”
  • “Return the correlation between columns A and B.”
  • “Give a 95% confidence interval for C2:C100.”

Correct formulas are:

=STDEV.S(B2:B80)
=CORREL(A:A, B:B)
=CONFIDENCE.T(0.05, STDEV.S(C2:C100), COUNT(C2:C100))

Sometimes it gets these right.
Sometimes it mixes things.
Sometimes it tries something clever.

That’s where the real testing begins.

🎯 What You’ll Learn by Testing

You’ll quickly see:

  • where small models shine
  • where they fail
  • how to prompt them better
  • how they think about nested formulas
  • what areas might benefit from micro-fine-tuning

Most importantly:

You’ll feel the gap between “big AI” and “small AI” narrowing — one formula at a time.

💬 Your Turn: Challenge the Model

Try these “stress test” prompts:

  • “Filter all rows where column B is ‘Active’ and return column D values.”
  • “Return the last non-empty cell in column E.”
  • “Calculate covariance between columns A and B.”
  • “Give a formula that returns TRUE if the text ‘error’ appears anywhere in D2.”

And this one, just for fun:

  • “Write a formula to rank the value in B2 among B2:B100.”

Watch how 6SigmaMind handles each case.

🛠️ Under the Hood (Simple Version)

6SigmaMind uses:

  • HuggingFaceTB/SmolLM2-1.7B
  • A lightweight Excel-specific system prompt
  • A Gradio interface
  • CPU-only inference on Hugging Face Spaces

Everything is open.
Everything is tiny.
Everything loads fast.

This isn’t enterprise AI — it’s exploratory AI.
And you’re invited to explore it.

🚀 Try It Now

👉 Test 6SigmaMind here:https://huggingface.co/spaces/benkemp/6SigmaMindv2
Just type a sentence and see what formula appears.

If you break it, even better.
That’s how we make Version 2 smarter.

🧭 What’s Next?

Next articles will explore:

  • fine-tuning small models for Excel
  • comparing small models (Gemma, Phi-3, Qwen, SmolLM2)
  • building datasets
  • pushing statistical knowledge into tiny AI
  • how to embed 6SigmaMind in an app

We’re only at the beginning.

Latest Articles