NumRs

High-Performance Numerical & Deep Learning Engine for Rust, Python, Node.js, R, and WASM.

[!WARNING] This project is currently in Alpha. APIs are subject to change and it is not yet recommended for mission-critical production use.

NumRs (Numeric Rust) is a next-generation numerical computing ecosystem. Built from scratch in Rust, it provides a unified, high-performance engine that powers bindings across the most popular programming languages. It combines the ease of use of NumPy/PyTorch with the safety and speed of safe Rust.

Whether you are training models in Python, deploying to the browser with WebAssembly, or building backend services in Node.js, NumRs guarantees the same behavior, same performance, and same API everywhere.

⚡ Key Features

• Universal Core: A single, robust Rust codebase powers all 5 bindings. Write logic once, run it anywhere.
• Hardware Acceleration: Auto-detects and uses AVX2/AVX-512 (x86) or Accelerate/Metal (Apple Silicon) for near-native speeds.
• Deep Learning Native: Built-in autograd engine (Reverse-mode AD) and a complete suite of Neural Network layers (Linear, Conv1d, LSTM, Transformer).
• Zero-Copy Architecture: Bindings share memory directly with the Rust core, eliminating serialization overhead for large tensors.
• Production Ready: Thread-safe parallel execution via Rayon, rigorous type checking, and guaranteed memory safety.

⚡ Performance

NumRs is engineered for speed. On an Apple M3 Max, NumRs achieves 2.42 Tops/s (Tera-operations per second) in matrix multiplication (FP32), utilizing Metal acceleration and SIMD optimizations.

Benchmark run via cargo run --bin numrs-bench --release

📄 View Full Benchmark Report

❤️ Contributors

Thank you to everyone who has contributed to NumRs!

💰 Sponsors

Support the development of NumRs!

🏗 Architecture

NumRs is designed as a layered ecosystem. The core engine handles the heavy lifting, while optimized bindings expose idiomatic APIs for each language.

graph TD
    NumRs[NumRs Core]
    
    subgraph Engine["Numerical Engine"]
        Array[Link: src/array]
        Ops[Link: src/ops]
        Backend[Link: src/backend]
    end
    
    subgraph ML["ML Framework"]
        Autograd[Link: src/autograd]
        NN[NN Modules]
    end
    
    subgraph Bindings["Universal Bindings"]
        JS[Node.js Binding]
        PY[Python Binding]
        WASM[WebAssembly]
        R[R Language]
        C[C ABI]
    end
    
    NumRs --> Engine
    NumRs --> ML
    Autograd --> Ops
    Ops --> Backend
    Ops --> Array
    Backend --> Array
    
    JS -.-> NumRs
    PY -.-> NumRs
    WASM -.-> NumRs
    R -.-> NumRs
    C -.-> NumRs

🌐 The Rosetta Stone

If it works here, it works everywhere.

Perform the same operations consistently across any language.

Matrix Multiplication with Autograd

🦀 Rust (Core)

use numrs::tensor::Tensor;

let a = Tensor::randn(&[2, 3], true);
let b = Tensor::randn(&[3, 2], true);
let c = a.matmul(&b);

c.backward();
println!("{:?}", a.grad());

🐍 Python

import numrs

# Drop-in replacement for standard workflows
a = numrs.randn([2, 3], requires_grad=True)
b = numrs.randn([3, 2], requires_grad=True)
c = a @ b

c.backward()
print(a.grad)

🟢 Node.js

import { Tensor } from '@numrs/node';

const a = Tensor.randn([2, 3], true);
const b = Tensor.randn([3, 2], true);
const c = a.matmul(b);

c.backward();
console.log(a.grad());

� ONNX Interoperability

NumRs models can be exported to standard ONNX format, allowing you to visualize them in Netron or deploy them with other runtimes.

🐍 Python Export

import numrs
# ... model training ...
dummy_input = numrs.randn([1, 10])
numrs.save_onnx(model, dummy_input, "model.onnx")

🟢 Node.js Export

import { Tensor } from '@numrs/node';
// ... model training ...
const dummyInput = Tensor.randn([1, 10]);
model.saveOnnx(dummyInput, "model.onnx");

🧠 Inference (Import)

NumRs can also load and run ONNX models (even users from PyTorch/TensorFlow).

Python

from numrs.onnx import OnnxModel

model = OnnxModel("model.onnx")
output = model.infer(dummy_input)

Node.js

import { OnnxModel } from '@numrs/node';

const model = OnnxModel.load("model.onnx");
const output = model.infer(dummyInput.data); // specific array

�🚀 Getting Started

Select your platform to get started:

🧭 Why NumRs Exists

Numerical computing shouldn’t be fragmented.

Today, numerical logic is rewritten for every platform: Python for research, JavaScript for the web, C++ for performance. Each rewrite introduces inconsistency, bugs, and friction.

NumRs exists to remove that fragmentation. A single Rust core defines the behavior of every operation, while each language provides only syntax, not semantics.

Write once. Run everywhere. Trust the math.

📄 License

NumRs is dual-licensed.

• Open Source: AGPL-3.0 (default). Free for open source, research, and internal use.
• Commercial: Contact us for a commercial license if you need to use NumRs in proprietary products without copyleft obligations.

Copyright (c) 2025 NumRs Contributors.