Vixana Download

Machine learning that builds itself.
On your own computer.

Drop in a spreadsheet and choose what to predict. Vixana builds, tests and combines many models for you, and your data never leaves the room.

Beat Amazon's AutoML on 82 of 103 tables and Microsoft's in our latest test · two registered benchmarks
curl -fsSL https://vixana.online/install.sh | sh

Give it a table. Get a model.

Open a file, choose the column you want to predict and how long Vixana may train. It works out the task and the score, trains and compares many models, and hands you the best as one predictor.

  • Experiments, models, predictions and Python notebooks in one window
  • Every run saved as a .vixana file you keep
  • No account with a cloud service, no upload, no bill per run
Download the app
The Vixana Studio showing a finished experiment: log loss 0.0459, a weighted ensemble of two models, trained in 48 seconds.
The Studio, on your machine

For data that should stay home

If your table holds patients, clients or anything you would rather not upload, Vixana trains where the data already is. You bring the table; it brings the models.

Doctors and clinics

Patient records stay in the clinic

You have
past visits and outcomes
Vixana predicts
who may miss an appointment, or be readmitted

A prediction supports your judgement; it does not replace it.

Researchers and students

A strong model without a cloud bill

You have
a study dataset
Vixana predicts
the outcome you study, as a baseline to beat or to report

Banks, insurers, accountants

Client data under your own roof

You have
past applications or claims
Vixana predicts
which ones deserve a second look

Shops and small businesses

No data team needed

You have
a sales spreadsheet
Vixana predicts
next month's demand, or which customers may leave

Engineers and factories

Works on the shop-floor PC

You have
sensor logs and inspection results
Vixana predicts
which part may fail, or which batch may not pass

Public services and charities

Sensitive cases, offline

You have
case records
Vixana predicts
which cases need help first
  1. Screen every candidate briefly
  2. Validate on rows it has not seen
  3. Stack a second level of models
  4. Select the blend that scores best
  5. Calibrate the probabilities

A data scientist's method, done for you

The careful steps an expert takes by hand, behind one button. Weak models are dropped early, every score comes from rows the model never saw, and the winners are blended into one predictor you can trust: when it says 80%, it is right about eight times in ten.

You set the time, from one minute to an hour. Vixana plans its work to fit it.

How Vixana works

Runs on the computer you already have

No graphics card? The CPU engine runs on any laptop: trees, neural networks and a pretrained model, stacked and blended.

Have an NVIDIA card, even a 4 GB one? The GPU engine adds Vixana's foundation model and sizes it to your card. That is the engine that won the test below.

Will it run on my computer?
82 of 103
tables where Vixana beat AutoGluon, the AutoML engine Amazon builds into SageMaker. 13 losses, 8 ties, over two registered benchmarks.
5 of 5
well-known AutoML systems beaten in the latest test, including FLAML from Microsoft Research.
+0.6–3.9%
typical (median) gain over AutoGluon, from stage to stage.

Beat Amazon's AutoML. And Microsoft's.

We tested Vixana against AutoGluon, the engine behind Amazon SageMaker Autopilot's ensemble mode, in two benchmarks whose rules were written before they ran: 114 tables, from a hundred rows to almost half a million. Vixana's GPU engine won 82 of the 103 tables that could be decided and lost 13.

In the latest test, on 30 fresh tables, it also beat FLAML from Microsoft Research and three more systems, with the same five minutes each on a 4 GB graphics card. No cloud, no rented servers.

The fine print: the ten largest tables, up to 470,000 rows, could not be decided on our 16 GB test laptop, because memory ran out for Vixana and AutoGluon alike; where they could be scored, they were near ties. The first benchmark's early stages used earlier versions of the GPU engine. The CPU engine did not reach the bar against AutoGluon in the latest test. The paper, coming soon, reports every number.

The systems it was tested against in the latest benchmark

  • AutoGluonAmazon · best quality
  • AutoGluonAmazon · extreme
  • FLAMLMicrosoft Research
  • H2O AutoMLH2O.ai
  • LightAutoMLSber AI Lab
  • MLJARMLJAR

Built for the machine you own

Your data stays home

Training runs on your laptop or workstation. Your tables, models and predictions never leave it, and there is no service to pay for.

Many models, one answer

Trees, neural networks and pretrained tabular models compete on the same folds. The best are stacked, blended and calibrated into one predictor.

From a CSV to a model

Drop in a file, choose the column to predict and how long to train. Vixana works out the task, the metric and the models.

Rules first, results second

We wrote down how Vixana would be judged before the first run, then published every result, the losses too.

  1. Rules sealedscores, time budgets and the win rule, fixed in advance
  2. New tables114 tables over two benchmarks; the latest 30 drawn after 585 known datasets were set aside
  3. Same machineevery system on the same laptop, with the same time budget
  4. All of it shown82 wins, 13 losses, 8 ties, and the tables we could not decide

A small GPU is enough

The GPU engine runs on a 4 GB consumer card. Without one, the CPU engine runs anywhere.

A model file you keep

Each run saves one .vixana file. Predict with it from the app, from the command line, or in the built-in Python notebooks.

Four kinds of model, each checked on rows it has not seen, joined into one predictor Gradient-boosted trees Neural networks Pretrained tabular model Foundation model · GPU CHECKED ON UNSEEN ROWS ONE PREDICTOR

“Many models, checked against data they have not seen, joined into one answer. On your own computer.”

Install

The desktop app sets itself up on first start, a few minutes once. The command line installs into your home folder and needs no administrator rights.

Linux, macOS, Windows

Terminal

One command, no administrator rights. Then run vixana.

curl -fsSL https://vixana.online/install.sh | sh

Current version: 6.2.0 (what's new). Each download is approved by the owner: sign in with Google, ask, and download once it is approved. Installers are not code-signed yet, so Windows SmartScreen may ask you to confirm. Read install.sh or install.ps1 before you run them.

Questions, answered

Does my data leave my computer?

No. The desktop app and the command line train on your machine. The network is used only to install the engine's packages the first time and to sign in with Google.

Is it as good as cloud AutoML?

In two registered benchmarks, Vixana's GPU engine beat AutoGluon, the engine Amazon uses in SageMaker Autopilot, on 82 of 103 tables, with the same time budget on the same small computer. Cloud services can rent bigger machines; the point of Vixana is that you don't have to, and that your data stays with you.

Is Vixana free?

Yes, for noncommercial use: personal projects, research, study and teaching, under the PolyForm Noncommercial License 1.0.0. For commercial use, contact the author.

Do I need a graphics card?

No. The CPU engine runs anywhere. The GPU engine needs an NVIDIA card with 4 GB or more; the docs show what it uses on each card.

Is there a Mac app?

Not yet. On a Mac, install the command line with the terminal command; it runs the CPU engine.

What can it predict?

A column in a table: a category (churn or not, which product, which diagnosis) or a number (price, demand, time). It reads CSV, TSV, Excel, Parquet, Feather and JSON files, and was tested on tables from 100 to 100,000 rows.

Why do I sign in with Google?

So the owner knows who is asking to download, and so the app knows who you are. Vixana receives your name, email address and picture, and nothing else from your Google account. See the privacy policy.

Your data. Your machine. Your model.

Free for personal projects, research, study and teaching. Install in one line, or download the app.

curl -fsSL https://vixana.online/install.sh | sh