What it does

Apache Mahout, once the go-to for distributed machine learning on Hadoop, has been reborn as a quantum computing stack. Its new flagship is Qumat, a Python library that abstracts quantum circuits across Qiskit, Cirq, and Amazon Braket backends. A companion module called QDP (Quantum Data Plane) uses Rust and CUDA to encode classical data into quantum states, with zero-copy tensor handoffs via DLPack to PyTorch, NumPy, and TensorFlow.

The interesting bit

The project is essentially two layers of abstraction: one for where you run quantum circuits (the backend-agnostic API), and one for how you get classical data onto a GPU in quantum-ready form. The DLPack zero-copy trick is the quiet workhorse here — it avoids the usual serialization tax when shuffling tensors between Python ML frameworks and Rust/CUDA kernels.

Key highlights

• Single API targets Qiskit, Cirq, and Amazon Braket; swap backends by changing a config dict
• QDP encodes classical vectors into quantum amplitude states via GPU kernels
• Rust core with PyO3 Python bindings; CUDA for GPU acceleration
• DLPack protocol enables zero-copy tensor sharing with PyTorch, NumPy, TensorFlow
• “Classic” Mahout (the Hadoop-era Java code) is now in maintenance mode

Caveats

• QDP’s PyTorch/NumPy/TensorFlow integration and additional encoders (angle, basis, multi-GPU) are still on the Q1 2026 roadmap, not yet shipped
• The project has 2,292 stars but appears to be in a transitional phase with much of the new Rust/CUDA stack still being hardened

Verdict

Worth watching if you’re building quantum ML pipelines and want to avoid vendor lock-in at the circuit level, or if you need GPU-accelerated state preparation without rewriting tensor plumbing. Skip it if you need production-ready quantum data encoding today — the QDP layer is still baking.