Don't import it. Build it.

Build your own ML framework: from tensors to systems.

Preview · Classroom release Spring 2027

An educational framework for building and optimizing ML: understand how PyTorch, TensorFlow, and JAX really work.

TinyTorch is usable today for self-paced learning and active course pilots. APIs, instructor packaging, and classroom workflows will continue to stabilize through the Spring 2027 classroom release. Self-study works today; courses adopting TinyTorch before then should expect to pilot it.

Start Building →

28,106 stars on GitHub: add yours and support free ML education

🔍 Click to Zoom

Tito CLI Companion & Environment Diagnostics

Build each piece: Tensors, autograd, attention. No magic imports.

Recreate history: Perceptron → CNN → TinyGPT → MLPerf & ChatGPT Serving.

Understand systems: Memory, compute, optimization trade-offs.

Debug anything: OOM, NaN, slow training, because you built it.

Recreate ML History

Walk through ML history by rebuilding its greatest breakthroughs with YOUR TinyTorch implementations. Click each milestone to see what you’ll build and how it shaped modern AI.

1958

The Perceptron

The first neural network, run forward with your layers

Input → Linear → Sigmoid → Output

1969

XOR Crisis

Minsky & Papert expose limits of single-layer networks

Input → Linear → Sigmoid → FAIL!

1986

MLP Revival

Backpropagation enables deep learning on TinyDigits

Images → Flatten → Linear → ... → Classes

1998

CNN Revolution

A LeNet-style convolutional network on handwritten digits

Images → Conv → Pool → ... → Classes

2017

Transformer Era

Train TinyGPT on Shakespeare, TinyCopilot code generation, and concept Q&A

Tokens → Embeddings → Causal Attention + MLP → Sampling

2018

MLPerf to Generative Serving

Production optimization & KV-cache generation speedup on TinyGPT

Profile → Quantize → Cache → Accelerate

2024 PRESENT

Custom Kernels

Verify your tiled, fused, and im2col kernels, then time them against C++ SIMD, Apple Metal MPS, and OpenAI Triton

Tiling → im2col → Fusion → Native reference kernels

Why Build Instead of Use?

"Building systems creates irreversible understanding."

Traditional ML Education

import torch
model = torch.nn.Linear(784, 10)
output = model(input)
# When this breaks, you're stuck

Problem: You can’t debug what you don’t understand.

TinyTorch: Build → Use → Reflect

# BUILD it yourself
class Linear:
    def forward(self, x):
        return x @ self.weight + self.bias

# USE it on real data
loss.backward()  # YOUR autograd