Vixana AutoML for your own models Download

Docs

What Vixana needs from your computer, what it does with your table, and the ways to use it.

What Vixana does

Give Vixana a table and choose the column you want to predict. Vixana works out whether that column is a category or a number and picks a fitting score. It then trains many kinds of model on the same splits of your data and tests each one on rows it did not learn from. The best are joined into one predictor.

The result is a model file you keep. Use it to predict new rows, today or a year from now, without sending your data anywhere.

The two engines

CPU engine
Vixana Core 5.7.1. Runs on any computer. Gradient-boosted trees, neural networks and a pretrained tabular model, stacked and blended.
GPU engine
Vixana G-final. The same pipeline plus Vixana's tabular foundation model, running on an NVIDIA graphics card. It is judged by the same rules as every other model in the run.

You don't have to choose a model size. When a GPU run starts, Vixana checks your card and sets itself up for it. Without a suitable card, use the CPU engine.

Will it run on my computer?

What Vixana needs
SystemWindows 10 or 11, 64-bit. Linux, 64-bit, as an AppImage, .deb or .rpm. On macOS, use the command line; it runs the CPU engine.
MemoryThe published results were measured with 8 GB of memory. Bigger tables need more.
ProcessorAny 64-bit processor. The published results used 2 threads; more cores let Vixana try more models in the same time.
Graphics cardNot needed for the CPU engine. The GPU engine needs an NVIDIA card with 4 GB of graphics memory or more.
Command linePython 3.11, 3.12 or 3.13, or uv, which can fetch one for you. No administrator rights.

Graphics cards

The GPU engine's foundation model learns from example rows of your table. How many it can read at once depends on your card's memory, so Vixana picks the model size for your card:

What the GPU engine uses on each card
Graphics memoryFoundation modelExample rows it reads
No NVIDIA cardNone; use the CPU engine—
Under 6 GB, such as 4 GBCompactUp to 8,000
6 GB to 16 GBMediumUp to 20,000
16 GB or moreLargeUp to 50,000

On a bigger table, the foundation model reads a sample of the rows as examples, and the other models still learn from every row. The published results were measured on a 4 GB card, the smallest size in the table.

Before a GPU run, close programs that use the graphics card heavily, such as games. Vixana treats a card that another program is busy with as not available.

Your data

  • Files: CSV, TSV, Excel, Parquet, Feather or JSON. One row per example.
  • What it predicts: one column. Either a category, such as churn or not, which product, which diagnosis, or a number, such as a price, demand or a time.
  • Size: tested on tables from 100 to 100,000 rows.
  • Where it stays: on your computer. Tables, models and predictions are kept in folders on the machine that runs Vixana.

How long a run takes

You choose. Vixana plans its work to fit the time you give it, and more time lets it train and compare more models. Very short times leave little room: on a big table, a one-minute run can take longer than asked.

The comparison on the home page gave every system five minutes per table. Vixana was also measured with one minute per table.

Ways to use it

The desktop app
The Studio: experiments, models, predictions and Python notebooks in one window. It sets itself up on first start, a few minutes once.
The command line
Install with one command, then run vixana for an interactive session, or vixana doctor to check your computer.
The model file
Every run saves one .vixana file. Predict with it from the app, the command line or a notebook.
A shared server
A Docker image runs Vixana for a team on one machine, with the same sign-in.

Downloads and accounts

Each download is approved by the owner. Sign in with Google, ask for approval, and download once it is approved; one approval covers one download. The install commands work the same way: they show a link and a code, and continue once you allow them.

Vixana is free for noncommercial use under the PolyForm Noncommercial License 1.0.0. See the privacy policy and the terms.

Go to downloads