The AI model
Totsum Data Explorer answers your questions with a large language model that runs locally, on your own machine, inside the application process. The app can work on a computer with no network connection. There is no account, no sync, no telemetry and no update check.
The model
Section titled “The model”Totsum uses Gemma 4 12B, a language model published by Google DeepMind under an open license, the Apache License 2.0. The installer places the model on your computer during setup, and the application itself has no download feature.
For the record, the file that Totsum runs is the version of that model prepared for local use by Unsloth (model card). The app shows the exact build on your machine under Help → About, together with the model’s license text.
How it runs
Section titled “How it runs”The model runs inside the application, on your own computer, on Windows and on macOS. It needs real hardware. See Hardware compatibility for the memory your machine must have.
What this means for your data
Section titled “What this means for your data”The model reads what the application hands it, namely your question and the schema and extracts of the files you opened. It computes its answers on your hardware, and those answers appear in the app. Exporting is how you take them out. AI-generated content in exports is marked, visibly and machine-readably.
Like every AI system, the model can be wrong. Totsum is built around that fact. Every answer comes with the query that produced it, and the follow-up report exists so that a reviewer can check the numbers before anyone relies on them.