Are local models good at maths and spreadsheets? What the benchmarks show, and a safe way to put a model near your numbers.
Short answer. Good at reasoning through maths problems, unreliable at raw arithmetic and spreadsheet edits unless they use tools. OpenAI’s gpt-oss card reports 92.5 per cent on AIME 2025 competition problems for the 120b model at high reasoning effort without tools and 97.9 per cent with a Python tool; at low effort the 20b model scores 37.1 per cent without tools and 57.5 with them. SpreadsheetBench, built from 912 real forum questions, found a substantial gap between models and humans on real spreadsheets. For office numbers, make the model write a formula or run code, and check totals against the source.
01
What was measured on competition maths
AIME is a set of competition mathematics problems. These are OpenAI’s own results for its open-weight gpt-oss models, from the model card’s evaluation table. “With tools” is OpenAI’s label; the card says the models are trained to use a Python tool that runs code, a browsing tool and developer-defined functions. Reasoning effort is a setting that lets the model think longer. Tools added 20 to 23 points at low effort and 5 to 7 points at high effort, so they helped most where the model thought least.
Model and reasoning effort
AIME 2025 without tools
AIME 2025 with tools
gpt-oss-120b, high effort
92.5 per cent
97.9 per cent
gpt-oss-120b, medium effort
80.0 per cent
91.6 per cent
gpt-oss-120b, low effort
50.4 per cent
72.9 per cent
gpt-oss-20b, high effort
91.7 per cent
98.7 per cent
gpt-oss-20b, medium effort
72.1 per cent
90.4 per cent
gpt-oss-20b, low effort
37.1 per cent
57.5 per cent
02
Spreadsheets are harder than sums
SpreadsheetBench was built from 912 real questions collected on online Excel forums, with spreadsheets that have several tables, irregular layouts and non-text elements, and several test files for each instruction. Its 2024 evaluation of many models found a substantial gap between the best models and human performance. Check the current leaderboard for newer models, but expect real workbooks to be harder than the tidy examples in a demo.
03
A safe way to use a model with numbers
Ask for a formula, not a value. A formula can be read, tested and reused; a typed-in number cannot be traced.
Let the model run code or a calculator tool where it is available, and have it show what it ran.
Recalculate in the spreadsheet and compare one total with the source document, every time.
Do not paste a whole workbook as text and ask for a corrected copy. Ask for the formulas to apply to the cells you name.
Keep the original file untouched, and keep a record of what was changed and by whom.
04
Where ANVE fits
ANVE’s agent loop allows up to seven tool steps per request, and a claim made without a successful tool call is challenged, which is the discipline above applied to a model’s statements. The Bills maker reads GST bills; this page gives no accuracy figure for it, so check totals against the printed bill. ANVE is pre-release and no sealed unit exists yet.
05
Questions people ask
Can a local model do our accounts?
Not unsupervised. It can draft formulas, explain a calculation and spot oddities, but each total should be checked by a person against the source. Treat its numbers as proposals.
Why do tools help so much?
A model writes text one piece at a time, so a long calculation can go wrong in the middle, while a calculator or code runs the same steps exactly. The table above shows the gain, largest where the model reasoned least.
Do bigger local models make fewer arithmetic mistakes?
Usually fewer, but the gpt-oss table shows that reasoning effort and tool use mattered more than size within that family. Test the model and the tools together.
06
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
