RUMUS
A deep learning framework with no Python in the training loop.
Written in native Rust with PyTorch-like ergonomics and its own WGPU compute backend. Open source under MIT and Apache, and the reason the rest of our compute layer exists.
SWUM TRAINS ITS MODELS WITH IT, NO PYTHON →202
STARS ON GITHUB
MIT
AND APACHE 2.0
AS OF SEPTEMBER 2026 · SOURCE: GITHUB →
WHY WE BUILT IT
The existing options forced a choice.
The Rust machine learning ecosystem made you pick between ergonomics and low level control. Frameworks reached for unsafe code and interior mutability to get around the borrow checker, missing fences stalled the GPU pipeline, and reference cycles in autograd held onto video memory long after it was needed.
We wanted the training loop to be fast, safe and legible at the same time, so we wrote one.
HOW IT WORKS
Train in Rust. Export to ONNX.
A full tensor data model over an op-driven autograd engine, a WGPU backend with per-resource fencing and mixed precision, fused optimizer kernels, INT8 quantization, an ONNX exporter, a just-in-time compiler with kernel fusion, and multi-GPU training. There's also a continuous batching inference server for serving.
Write a model natively, export it to ONNX, and serve it through OmniNode with a receipt on every answer.
THE SPECIFICS
