Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/MilesONerd/neurenix/llms.txt

Use this file to discover all available pages before exploring further.

​
Introduction to Neurenix

Neurenix is an artificial intelligence framework optimized for embedded devices (Edge AI), with support for multiple GPUs and distributed clusters. The framework specializes in AI agents, with native support for multi-agent systems, reinforcement learning, and autonomous AI.
Neurenix is designed to empower intelligent futures, one edge at a time. It brings state-of-the-art AI capabilities to resource-constrained environments without sacrificing performance.

​
What Makes Neurenix Different?

Neurenix stands out in the AI framework landscape through several key innovations:

Edge-First Design

Optimized for embedded devices and edge computing, making AI accessible on resource-constrained hardware

Hot-Swappable Backends

Runtime device switching with automatic hardware detection through the Genesis system

Multi-Agent Native

Built-in support for multi-agent systems, reinforcement learning, and autonomous agents

Universal Hardware

Support for CPU, CUDA, ROCm, WebGPU, TPU, NPU, Vulkan, OpenCL, oneAPI, DirectML, TensorRT, ARM, and more

​
Core Features

​
Hardware Acceleration

Neurenix provides comprehensive hardware acceleration support across diverse platforms:
  • GPU Backends: CUDA, ROCm, WebGPU, Vulkan, OpenCL, DirectML
  • AI Accelerators: TPU, NPU, Tensor Cores, GraphCore IPU, FPGA
  • Optimized Libraries: oneDNN, MKL-DNN, TensorRT
  • ARM Architecture: Compute Library, Ethos-U/NPU, Neon SIMD, SVE
  • WebAssembly: SIMD and WASI-NN for browser-based execution

​
AI Capabilities

1

Agent-Based AI

Native support for single and multi-agent systems with reactive, deliberative, and hybrid architectures. Includes communication protocols, coordination mechanisms, and multi-agent learning algorithms.
2

Advanced Learning

Meta-learning (MAML, Reptile, Prototypical Networks), transfer learning, continual learning, federated learning, and zero-shot learning capabilities.
3

Neuro-Symbolic AI

Hybrid models combining neural networks with symbolic reasoning, differentiable logic, and knowledge distillation.
4

Specialized Networks

Graph Neural Networks (GNN), fuzzy logic systems, multi-scale architectures, and quantum computing integration.

​
Model Optimization

Neurenix includes powerful optimization capabilities: 
from neurenix import Tensor, Device, DeviceType
from neurenix.quantization import quantize_model, QuantizationConfig

# Automatic quantization support
config = QuantizationConfig(
    precision="int8",  # INT8, FP16, or FP8
    calibration_data=calibration_set,
    pruning_enabled=True
)

quantized_model = quantize_model(model, config)
Quantization can reduce model size by up to 4x while maintaining accuracy, making deployment on edge devices feasible.
  • Hyperparameter search (Grid, Random, Bayesian, Evolutionary)
  • Neural architecture search (NAS) for automated model design
  • Model selection and pipeline optimization
  • Neuroevolution with NEAT, HyperNEAT, and CMA-ES

​
Distributed Training

Scale your training across multiple nodes and devices:
  • MPI for high-performance computing clusters
  • Horovod for distributed deep learning
  • DeepSpeed for large-scale model training
  • Asynchronous training with automatic checkpointing and resume

​
Production-Ready Features

Model Serving

RESTful, WebSocket, and gRPC APIs for serving models in production

ONNX Support

Import/export models to ONNX for interoperability with other frameworks

Explainability

SHAP, LIME, feature importance, and activation visualization

Containerization

Docker and Kubernetes support for cloud-native deployments

​
Architecture Overview

Neurenix is built on a layered architecture that separates concerns and enables flexibility: 
┌─────────────────────────────────────────────────────────┐
│                   High-Level APIs                       │
│  (Agents, AutoML, Federated Learning, Quantum, etc.)   │
└─────────────────────────────────────────────────────────┘
                          |
┌─────────────────────────────────────────────────────────┐
│                  Neural Network Layer                   │
│        (Module, Linear, Conv2d, LSTM, etc.)            │
└─────────────────────────────────────────────────────────┘
                          |
┌─────────────────────────────────────────────────────────┐
│                    Tensor Operations                    │
│         (Tensor, automatic differentiation)            │
└─────────────────────────────────────────────────────────┘
                          |
┌─────────────────────────────────────────────────────────┐
│                   Device Abstraction                    │
│          (DeviceManager, Genesis system)               │
└─────────────────────────────────────────────────────────┘
                          |
┌─────────────────────────────────────────────────────────┐
│                  Hardware Backends                      │
│  (CUDA, ROCm, Vulkan, TPU, NPU, WebGPU, etc.)         │
└─────────────────────────────────────────────────────────┘
                          |
┌─────────────────────────────────────────────────────────┐
│                    Phynexus Engine                      │
│        (Rust/C++/Go optimized core operations)         │
└─────────────────────────────────────────────────────────┘

​
Key Components

Tensor System: The foundation of Neurenix, providing n-dimensional arrays with automatic differentiation and device-agnostic operations. Device Management: Hot-swappable backend functionality with runtime device switching. The Genesis system automatically detects and selects optimal hardware. Neural Network Modules: Composable building blocks for creating models, including Linear, Conv2d, LSTM, and more specialized layers. Agent Framework: Complete infrastructure for building autonomous agents, including environments, policies, value functions, and multi-agent coordination. Optimization: Advanced optimizers (SGD, Adam, etc.) with support for distributed training and gradient accumulation.

​
Use Cases

Neurenix excels in scenarios requiring:
1

Edge AI Deployment

Deploy models on IoT devices, embedded systems, and edge servers with minimal resource footprint
2

Multi-Agent Systems

Build collaborative AI systems for robotics, game AI, autonomous vehicles, and swarm intelligence
3

Heterogeneous Computing

Leverage diverse hardware accelerators in a single application with automatic device selection
4

Research & Experimentation

Explore cutting-edge AI techniques like neuro-symbolic learning, quantum ML, and meta-learning

​
Community & Support

Join the Neurenix community:
Neurenix is currently in Alpha stage (v2.0.1). APIs may change between versions. Use in production with caution.

​
Next Steps

Installation

Get Neurenix installed on your system

Quickstart

Build your first model in 5 minutes

Build docs developers (and LLMs) love

Get started for freeTalk to us