jax4j / Developer workbench

Numerical Java.
Ready to compose.

A tensor API, automatic differentiation and JIT compilation in one Java 25 toolkit.

FusedFunction.javaJAVA 25

var function = Jit.compile(
    x -> x.mul(x).add(3).tanh()
);

try (Tensor output = function.apply(input)) {
    output.blockUntilReady();
}

Capture on the first call. Replay with new runtime values.

$ JAX4J_BACKEND=cpu java …

01 /

Think in tensors.

A fluent Java API for tensor operations, linear algebra and neural networks.

02 /

Differentiate the function.

Gradients, JVPs, VJPs, Jacobians and Hessians. Compose transforms with JIT and batching.

03 /

Choose your compute.

CPU execution in core. Vulkan and wgpu validated in 0.2.1; optional CUDA, ROCm and SIMD have separate validation limits.

Start with the core

Add a dependency.
Bring your function.

The core includes CPU kernels, graph capture, autodiff, linear algebra, neural network utilities and Adam. Native engines use Java’s Foreign Function & Memory API.

Setup and native access →

pom.xmlMAVEN

<dependency>
  <groupId>com.nablatensor</groupId>
  <artifactId>jax4j-core</artifactId>
  <version>0.2.1</version>
</dependency>

Add an optional engine for GPU execution.

Built on jax4j / Quant

Put the compute
to work in finance.

Quant 0.4.0 owns the financial models: products, Monte Carlo valuation, calibration, risk and CVA. jax4j supplies the tensors, differentiation and compute backends underneath.

PricingRiskCVA

Explore Quant ↗Open QuantLab →