Local AI models on your phone can do everyday math, explain concepts and set up word problems well, but they make arithmetic mistakes, especially with long numbers, many steps or big lists of figures. Turning on thinking mode, where the model works through the problem step by step, catches many of those errors. For any number that matters, check the final calculation with a calculator.
The trick is to use the AI for what it’s good at, the reasoning and setup, and let a calculator do the arithmetic.
Why are language models bad at arithmetic? #
A language model doesn’t calculate. It predicts the next piece of text based on patterns it learned. For “2 + 2 =”, the pattern is so common that it gets it right. For “4,738 × 219 =”, it has to produce each digit by prediction, and it can slip on any of them.
Text is also split into tokens, chunks that don’t line up neatly with digits. A long number might be split into odd pieces, which makes digit-by-digit work harder.
Smaller models, the kind that fit on a phone, have less capacity for this. So the gap between “explain this” and “compute this exactly” is wider on-device than with large cloud models.
What math can a local model handle? #
| Task | How reliable | Tip |
|---|---|---|
| Explaining a concept (what’s a derivative?) | Good | Ask for an example |
| Setting up a word problem | Good | Ask it to define the variables |
| Simple percentages and tips | Good, usually | Check the final number |
| Unit conversions | Fair | Verify the conversion factor |
| Algebra step by step | Fair to good | Use thinking mode |
| Multi-digit multiplication | Weak | Use a calculator |
| Adding up a long list of numbers | Weak | Use a spreadsheet or calculator |
| Statistics on pasted data | Weak | Ask for the formula, compute yourself |
| Writing a spreadsheet formula | Good | Test it on known values |
| Logic puzzles | Good with thinking mode | Check the reasoning |
How does thinking mode help with math? #
Thinking mode makes the model write out its reasoning before it answers. Working through intermediate steps gives it room to catch its own mistakes, much like showing your work on paper.
A classic example: a bat and a ball cost $1.10 together, and the bat costs $1.00 more than the ball. How much is the ball? The quick, intuitive answer is 10 cents, which is wrong. With reasoning written out, the model sets up ball + (ball + 1.00) = 1.10 and gets 5 cents.
In Personal LLM, thinking mode shows that working in a collapsible Reasoning panel above the answer, so you can check each step and fold it away when you’re done. It’s off by default because it makes answers slower, and most questions don’t need it. Turn it on with the Thinking preset for math, logic and anything you want to verify. The catalog’s Qwen 3.5 9B is the strongest option for this if your phone has 6 GB of RAM or more.
How to get correct math answers from a local AI #
- Turn on thinking mode for anything with more than one step.
- Use the Precise preset for calculations. Lower temperature means less randomness in the answer.
- Ask it to show its work. “Show each step” makes errors visible.
- Split big problems. Ask for step one, check it, then ask for step two.
- Let it set up, you compute. “Write the formula I need to calculate the monthly payment” is more reliable than “what’s my monthly payment?”
- Ask it to check itself. “Plug your answer back into the original problem and verify it” catches some mistakes.
- Verify with a calculator. Always, for money, medicine, measurements or grades.
- Use a bigger model for hard problems. You can switch to a larger model mid-conversation.
Working with numbers and data #
You can paste numbers into a chat as text, like a list of expenses or a small table, and ask the AI to organize them. It’s good at:
- Grouping expenses into categories
- Turning messy notes into a clean table
- Writing the spreadsheet formula to total a column
- Explaining what a trend in the numbers might mean
It’s not good at adding up 40 numbers exactly. Paste them into a spreadsheet for that, and use the AI to write the formulas. If those numbers are your bank transactions, keep them on your phone: our guide on sharing financial info with AI explains why.
Can I use local AI for homework? #
It’s a good tutor and a poor answer key. Use it to:
- Explain a concept a different way when the textbook doesn’t click
- Walk through a similar example step by step
- Check your reasoning (“Where did I go wrong here?”)
- Generate practice problems
Don’t copy its final answers without checking. It can be confidently wrong, and you won’t learn the method. A system prompt helps turn it into a tutor; see our system prompt examples for one that asks guiding questions instead of giving answers.
Why does thinking mode make it slower? #
The model writes out its reasoning first, often several times more text than the final answer, and every word takes time and battery. On a phone, that can mean a long wait. Use it when accuracy matters and skip it for quick questions. Our guide to speeding up local AI has more on the trade-off.
Math mistakes are one form of a wider problem, where models produce confident text that isn’t true. Why AI makes things up covers the rest.
Frequently asked questions #
Is AI good at math? #
Large AI models have become fairly good at math, especially with step-by-step reasoning. Small on-device models are decent at explaining and setting up problems but still make arithmetic errors. Check final numbers with a calculator.
Why does AI get simple math wrong? #
Language models predict text rather than calculate. Common sums are well represented in their training data, but unusual numbers have to be produced digit by digit through prediction, and any digit can come out wrong.
Does thinking mode make AI better at math? #
Usually, yes. Writing out intermediate steps lets the model catch errors it would make answering in one go. It’s slower, so use it for problems with several steps.
Can offline AI solve word problems? #
Often, yes. Small models are good at turning a word problem into equations. Use thinking mode, ask it to show each step, and check the arithmetic yourself.
Should I trust AI for money calculations? #
Use it to understand and set up the calculation, then do the math with a calculator or spreadsheet. Don’t rely on an AI’s arithmetic for loans, taxes or budgets without checking.