WebTabPFN

Built with PriorLabs TabPFN

WebTabPFN is an independent, unofficial browser port and is not affiliated with or endorsed by Prior Labs.

WebTabPFN brings TabPFN v2 classification and regression to the browser from one classic script. It downloads prepared INT4 or INT8 weights only when needed, keeps them in the browser cache, and runs them through an explicitly selected WebGPU or WASM backend. Input data never leaves the browser.

• Runs locally: training data and predictions stay in the browser.
• Hardware aware: choose WebGPU or WASM explicitly.
• Lazy and cached: weights download on demand and are reused across visits.

/not-a-feature/WebTabPFN /package/webtabpfn

Try it

Paste a CSV with a header row. Pick the target column, then leave that target blank (or use ?) for every row you want WebTabPFN to predict. Feature columns must be numeric; class labels may be words or numbers.

CSV playground

● Data stays in your browser

sepal_length,sepal_width,petal_length,petal_width,species 5.1,3.5,1.4,0.2,setosa 4.9,3.0,1.4,0.2,setosa 5.0,3.4,1.5,0.2,setosa 6.4,3.2,4.5,1.5,versicolor 6.9,3.1,4.9,1.5,versicolor 5.7,2.8,4.1,1.3,versicolor 6.5,3.0,5.8,2.2,virginica 7.6,3.0,6.6,2.1,virginica 7.2,3.6,6.1,2.5,virginica 5.0,3.6,1.4,0.2,? 6.1,2.9,4.7,1.4,? 7.2,3.2,6.0,1.8,?

12 rows · 5 columns Restore example

Target column Model INT4 · 4.9 MB · fastest WebGPU INT8 · 7.7 MB · recommended for WASM FP32 · 28.1 MB · reference Predict missing labels→

INT4 is the smallest download and the default choice for WebGPU.

Ready — 9 training rows and 3 rows to predict.

The first prediction downloads the selected weights. Later runs reuse the cached model, so they start much faster.

WebTabPFN can also be used for regression by loading it with task: "regression"; the fit() and predict() workflow stays the same.

Use it in a page

Pin an exact package version in production. Loading the script itself does not fetch a model; weights are downloaded lazily by load().

<script src="https://cdn.jsdelivr.net/npm/webtabpfn@0.2.0/src/webtabpfn.js"></script>
<script>
  const classifier = await WebTabPFN.load({
    task: "classification",
    backend: "wasm",
    precision: "int8",
  })
  classifier.fit(
    [[0.1, "low"], [0.2, "low"], [1.1, "high"], [0.9, "high"]],
    ["control", "control", "case", "case"],
  )
  const predictions = await classifier.predict([[0.8, "high"]])
</script>

load() requires a task ("classification" or "regression"), a backend ("webgpu" or "wasm"), and a precision ("int4" or "int8"). Numeric, string, boolean, and missing feature values are supported.

Speed versus accuracy

The browser benchmarks cover eight deterministic classification datasets. INT4 on WebGPU offers the strongest deployable latency/accuracy trade-off in the current measurements; exact performance depends on the browser and hardware.

License

WebTabPFN’s original source code is available under Apache-2.0. The bundled TabPFN weights use the Prior Labs License v1.2, and ONNX Runtime Web is licensed by Microsoft under the MIT License. WebTabPFN is an independent, unofficial browser port and is not affiliated with or endorsed by Prior Labs.