8,100+ classes • 139 modules • Pure C# • Open Source

The complete AI platform
for .NET developers.

Train, serve, and monetize AI models entirely in C#. From neural networks toRAG pipelines,computer vision toreinforcement learning — with custom GPU kernels that areup to 500x faster than PyTorch. No Python. No wrappers. No compromises.

Get Started Free View on GitHub

$dotnet add package AiDotNet

x GPU Speedup vs PyTorch

Public Classes

AI/ML Modules

Python Dependencies

.NET 10 + .NET Framework 4.7.1BSL 1.1 Licensed

Faster than PyTorch. On any GPU.

AiDotNet.Tensors ships hand-tuned GPU kernels via CUDA, OpenCL, HIP, Metal, Vulkan, and WebGPU — no CUDA lock-in, no third-party GPU libraries. Your code runs on NVIDIA, AMD, Intel, and Apple GPUs with zero configuration.

500x

GEMM (2048x2048)

Matrix multiplication speedup vs CPU

200x

Conv2D

Convolution speedup for vision models

100x

GEMV

Matrix-vector multiply speedup

40x

MaxPool2D

Pooling layer speedup

Adaptive thresholds automatically route small operations to CPU and large operations to GPU for optimal performance.

NVIDIA (CUDA)AMD (OpenCL)Intel (OpenCL)CPU (SIMD/AVX-512)

Every AI domain. One NuGet package.

139 modules covering neural networks, classical ML, computer vision, audio, NLP, reinforcement learning, model serving, and more.

Any GPU

GPU Acceleration

Custom GPU kernels for NVIDIA, AMD, Intel, and Apple GPUs via CUDA, OpenCL, HIP, Metal, Vulkan, and WebGPU. Up to 500x GEMM speedup. Faster than PyTorch.

160+

Neural Networks

CNN, RNN, Transformer, GAN, VAE, GNN, Diffusion, NeRF and more. 160 layer types covering every major architecture.

115+

Computer Vision

YOLO v8-11, DETR, RT-DETR, SAM/SAM2, segmentation, OCR, text detection. Full detection-to-recognition pipeline.

75+

Diffusion & Generation

40+ image diffusion models (SD, SDXL, SD3, Flux, ControlNet), 29 video generation models, 7 audio diffusion models.

155+

Classical ML

50 classifiers, 57 regressors, 48 clustering algorithms. Complete traditional ML toolkit with AutoML and NAS.

250+

Audio & Speech

Speech recognition, text-to-speech, music generation, audio classification. 250+ implementations across the full audio pipeline.

165+

Vision-Language Models

LLaVA, InternVL, CLIP, DINOv2/v3, Florence2, Qwen-VL. Instruction-tuned VLMs and vision encoders.

50+ agents

Reinforcement Learning

DQN, PPO, SAC, A3C, TD3, MuZero, Decision Transformer, MADDPG. 50+ agents for games, robotics and optimization.

Full stack

Model Serving & API

Production REST API with AiDotNet.Serving. Model encryption, license management, API key auth, Stripe billing. Deploy your AI as a product.

50+ adapters

LoRA Fine-tuning

LoRA, QLoRA, DoRA, AdaLoRA, VeRA, MoRA and more adapters plus training strategies. Fine-tune models with minimal compute.

180+

Federated & Distributed

FedAvg, FedProx, scaffold, and 30+ distributed training strategies. Train across hospitals, banks, or edge devices while keeping data private.

30+ models

NER & NLP

BERT, BioBERT, FinBERT, LegalBERT, DeBERTa and 30+ named entity recognition models. BiLSTM-CRF, span-based and transformer NER.

Browse all 20+ categories

One API for everything

AiModelBuilder is a single fluent API with 130+ configuration methods. Simple for beginners, fully customizable for experts.

var result = await new AiModelBuilder<double, double[], double>()
    .ConfigureModel(new RandomForestClassifier<double>(nEstimators: 100))
    .ConfigurePreprocessing()
    .BuildAsync(features, labels);

var prediction = result.Predict(newSample); // That's it.

var result = await new AiModelBuilder<double, double[], double>()
    .ConfigureModel(new NeuralNetwork<double>(
        inputSize: 4,
        hiddenSize: 16,
        outputSize: 3))
    .ConfigureOptimizer(new AdamOptimizer<double>())
    .ConfigurePreprocessing()
    .BuildAsync(features, labels);

var prediction = result.Predict(newSample);

var result = await new AiModelBuilder<float, float[], float>()
    .ConfigureModel(new YoloV11<float>(
        modelPath: "yolov11n.onnx",
        classes: CocoClasses.Names))
    .ConfigurePreprocessing()
    .BuildAsync(images, labels);

var detections = result.Predict(newImage);

var result = await new AiModelBuilder<double, double[], double>()
    .ConfigureModel(new RandomForestClassifier<double>())
    .ConfigurePreprocessing()
    .ConfigureQuantization(quantConfig)
    .BuildAsync(features, labels);

// Deploy as REST API
result.ServeAsApi(app, "/api/predict");

Built for your industry

Purpose-built AI pipelines for regulated industries. Complete workflows from data ingestion to production deployment.

Financial Services

FinBERT sentiment, DeepAR forecasting, GARCH volatility, fraud detection. Native .NET performance for trading systems.

FinBERTChronosIsolationForest

Healthcare

Medical imaging with MedSAM, clinical NLP with BioBERT, and federated learning for HIPAA-compliant cross-hospital training.

MedSAMBioBERTFedAvg

Document AI

OCR, layout analysis, key-value extraction from invoices, contracts, and medical records. Full pipeline in pure .NET.

PaddleOCRLayoutLMv3Donut

How we compare

AiDotNet is the only .NET platform that covers training, serving, and monetization. No compromises.

Feature comparison between AiDotNet, ML.NET, TorchSharp, and PyTorch
Feature	AiDotNet	ML.NET	TorchSharp	PyTorch
Language	Pure C#	C# + ONNX	C# wrapper	Python
Neural Network Layers	160+	~10	50+	100+
Classical ML Algorithms	155+	~30	None	None
Computer Vision Models	115+	Limited	Via LibTorch	Via torchvision
Vision-Language Models	165+	None	None	Via HF
Diffusion Models	75+	None	None	Via diffusers
Audio / Speech / TTS	250+	None	Limited	Via torchaudio
Reinforcement Learning	50+	None	Manual	Via SB3
GPU Support	Any GPU (6 backends)	ONNX only	CUDA only	CUDA only
Federated Learning	180+	None	None	Via Flower
Model Serving	Built-in	Manual	None	TorchServe
Runtime Dependency	None	None	~700MB LibTorch	Python runtime

Why not TorchSharp?

~700MB LibTorch runtime dependency
Slow startup loading PyTorch runtime
Array copying between .NET and LibTorch
No classical ML, no AutoML, no serving

Why not ML.NET?

Only ~10 neural network architectures
No computer vision, audio, VLMs, or RL
No GPU acceleration or custom kernels
No model serving or monetization tools

Why not Python?

GIL limits real concurrency
Separate runtime in production stack
No native .NET type safety or AOT
Separate team/skills from your .NET devs

Native .NET performance

No Python runtime. No C++ interop. Pure .NET speed with hardware-level optimizations.

SIMD

AVX2/AVX-512 vectorized tensor operations

Span<T>

Zero-copy memory operations throughout

AOT

Ahead-of-time compilation support

6 GPU APIs

CUDA, OpenCL, HIP, Metal, Vulkan, WebGPU

From training to production

AiDotNet covers the entire lifecycle. Train models, deploy as APIs, and manage customers — all in .NET.

AiDotNet

The core library. 8,100+ classes for training, inference, and experimentation. Install from NuGet and start building.

TrainingInferenceAiModelBuilder

AiDotNet.Tensors

Hand-tuned GPU kernels via P/Invoke. CUDA, OpenCL, HIP, Metal, Vulkan, and WebGPU backends. Adaptive CPU/GPU routing for optimal performance.

GPU Kernels6 BackendsAny GPU

AiDotNet.Serving

Deploy models as REST APIs. Encrypted model protection, API key auth, license management, and Stripe billing built in.

REST APIEncryptionStripe

Ready to build AI in .NET?

Install from NuGet and start training models in minutes. Training and inference are always free.

Get Started Free View Pricing

The complete AI platformfor .NET developers.