A high-performance Deep Learning framework with a Rust core, native to the Node.js ecosystem.

Forge your own neural networks — no Python, no C/C++ toolchain, no heavyweight runtime.

English · Español

Intellivium (formerly NeuroForge) is a deep learning framework whose numerical core is written entirely in Rust and exposed to JavaScript/TypeScript through N-API. The goal is the power of a native ML engine with the ergonomics of the npm ecosystem: npm install, import, and train — with precompiled binaries per platform and no C/C++ source, no Python, and no embedded VM.

It draws inspiration from PyTorch, TensorFlow and Flux.jl, but makes a deliberate engineering choice: one native language (Rust) for the engine, TypeScript for the public API, and Zig reserved strictly for future hot kernels.

Why not Julia / C++? Julia can be embedded, but it drags its full runtime (VM + LLVM + stdlib, hundreds of MB) plus JIT warm-up — impractical to ship over npm. C++ is unnecessary: Rust delivers the same low-level control and SIMD without the toolchain pain. See Architecture.

v0.11.0 · on npm. The engine is tested and trains real models. It's still pre-1.0, so the API may evolve — and the grand vision further down is a roadmap, not a current claim.

Available today ✅

• Reverse-mode automatic differentiation (Wengert tape, no Rc<RefCell>).
• Ops: matmul, broadcasted bias add, relu, sigmoid, tanh, MSE.
• Dense layers with He init, sequential Model, SGD & Adam optimizers, MSE / BCE / categorical cross-entropy losses, softmax for multi-class output, extra activations (LeakyReLU / ELU / GELU) and MAE / Huber losses, tensor ops & views (add/sub/mul, matmul, reductions, one-hot, reshape/transpose/slice), plus Dropout, LayerNorm, BatchNorm and Embedding layers, mini-batch training with Dataset/DataLoader, validation + early stopping + checkpoints, custom Dataset, weight export/import, gradient clipping, LR decay, and model save/load.
• N-API bindings + typed TypeScript API.
• Validated end-to-end on the XOR problem (non-linear): loss 0.247 → 0.0002.

npm install intellivium

import { tensor, dense, Model } from "intellivium";

// XOR — the classic non-linear sanity check
const X = tensor([[0, 0], [0, 1], [1, 0], [1, 1]]);
const y = tensor([[0],    [1],    [1],    [0]]);

const model = new Model([
  dense(2, 8, "tanh"),
  dense(8, 1, "sigmoid"),
]);

const history = await model.train(X, y, {
  epochs: 1500,
  lr: 0.05,
  optimizer: "adam",
  loss: "bce",
});
console.log("final loss:", history.at(-1));

const pred = model.predict(X);
console.log(pred.toArray()); // ≈ [[0], [1], [1], [0]]

The autograd graph lives entirely in Rust. Only flat tensors (Float64Array + shape) and high-level operations cross the FFI boundary — the graph is never marshalled per-op.

flowchart TD
    subgraph JS["TypeScript / JavaScript"]
        A["Public API<br/>tensor · dense · Model"]
    end
    subgraph NAPI["neuroforge-napi (cdylib)"]
        B["N-API bridge<br/>flat tensors in / out"]
    end
    subgraph CORE["neuroforge-core (pure Rust)"]
        C["Autograd tape<br/>matmul · add · relu · sigmoid · tanh · mse"]
        D["nn: Dense · Model · SGD/Adam"]
        E["(future) Zig SIMD kernels"]
    end
    A -->|Float64Array + shape| B
    B --> C
    D --> C
    C -.->|hot path| E

Layout

Intellivium/
├── crates/
│   ├── neuroforge-core/    # pure-Rust engine: autograd + nn  ← works today
│   │   ├── src/tape.rs     #   reverse-mode AD (the heart)
│   │   ├── src/nn.rs       #   Dense, Model, train/predict
│   │   └── examples/xor.rs #   `cargo run` proof
│   └── neuroforge-napi/    # N-API bindings → .node
├── src/                    # public TypeScript API
└── examples/               # xor.mjs (Node)

Requirements: Rust (rustup), Node.js 18+, and @napi-rs/cli (already a dev dependency).

# 1. test the Rust engine alone (no Node needed)
cargo run --release -p neuroforge-core --example xor

# 2. build everything (native .node + TypeScript)
npm install
npm run build

# 3. run the Node example
npm test

The engine is the foundation. Everything below is the long-term plan, phase by phase — honest about what already ships versus what's still ahead.

Legend: ✅ Complete · 🟡 Partial · 🔴 Planned

A stable, reliable engine.

Tensors

• Data types (dtype system — currently f32 only)
• Broadcasting
• Basic operations
• MatMul
• Shape checking
• Tensor views

Autograd

• Reverse mode
• Wengert tape
• Gradients
• Computation graph
• Automatic graph release

Neural networks

• Dense
• Sequential
• He initialization
• ReLU
• Sigmoid
• Tanh
• Softmax
• LeakyReLU
• ELU
• GELU

Loss functions

• MSE
• BCE
• Cross-Entropy (categorical)
• MAE (L1)
• Huber

Optimizers

• SGD
• Adam

API

• TypeScript
• N-API
• npm

Make training a model comfortable.

Dataset

• Dataset
• TensorDataset
• Custom Dataset

DataLoader

• Mini-batch
• Shuffle
• Batch iterator

Training loop

• Validation
• Early stopping
• Checkpoints
• Gradient clipping
• Learning-rate scheduler

Serialization

• save()
• load()
• exportWeights()
• importWeights()

More building blocks.

Layers

• Dropout
• BatchNorm
• LayerNorm
• Embedding
• Flatten
• Reshape

CNN

• Conv1D
• Conv2D
• Conv3D

Pooling

• MaxPool
• AvgPool
• Global pool

Recurrent

• RNN
• LSTM
• GRU

Before adding modern AI.

SIMD

• Zig SIMD
• Optimized MatMul
• Optimized Conv

Memory

• Arena allocator
• Buffer pool
• Zero copy
• Tensor pool

Parallelism

• Rayon
• Multi-thread

Benchmark

• Benchmarks
• Profiler

Where Transformers finally appear.

Attention

• Self-attention
• Multi-head attention
• Positional encoding
• Rotary embeddings

Transformers

• Encoder
• Decoder
• GPT
• BERT

Vision

• Vision Transformer
• ConvNeXt

• Autoencoder
• VAE
• GAN
• Diffusion

• Replay buffer
• DQN
• PPO
• SAC
• Actor-Critic

Scientific computing.

Linear algebra

• LU
• QR
• SVD
• Eigen

Numerical

• Optimization
• ODE
• Root finding

Statistics

• Monte Carlo
• Distributions

Data

• Lazy loading
• Streaming
• Hot cache
• Dataset cache

Formats

• Parquet
• Arrow
• CSV
• JSON

Engine

• Memory mapping
• Prefetch
• Columnar storage

GPU

• CUDA
• ROCm
• Metal
• Vulkan

Mixed precision

• FP16
• BF16

Quantization

• INT8
• INT4

• Inference engine
• Batch inference
• ONNX
• Serving
• HTTP
• gRPC

Tooling, not AI.

Visualization

• Dashboard
• Tensor inspector
• Training graphs

Extensions

• Plugins
• Custom layers
• Custom optimizers

Model Hub

• Models
• Datasets

The most ambitious vision.

Distributed

• Multi-GPU
• Multi-node

Compiler

• Graph optimizer
• Kernel fusion

AutoML

• NAS
• Hyperparameter search

Issues, ideas and pull requests are welcome. For substantial changes, open an issue first to discuss direction. Please keep the engine (neuroforge-core) free of binding/runtime concerns — that separation is intentional.

Apache License 2.0. You may use, modify, and distribute this software under the terms of the Apache 2.0 license, including a patent grant. Copyright © 2026 Brashkie.