Machine Learning Systems

TWO-VOLUME TEXTBOOK

Machine Learning
Systems.

The physics of AI engineering.

A rigorous, principles-first treatment of how ML systems are built, optimized, and deployed — from a single machine to fleet-scale infrastructure.

Harvard University · MIT Press 2026

Actively maintained Last updated September 2026 Release notes

Volume I cover

Volume I

Introduction to Machine Learning Systems

HTML PDF

Volume II cover

Volume II

Scaling Machine Learning Systems

HTML PDF

Explore the Curriculum
↓

A complete curriculum for AI engineering.

Choose a path: read the books, explore trade-offs in labs, build the internals with TinyTorch, model constraints with MLSys·im, deploy on real hardware, practice with StaffML, or adopt the full course with the Blueprint.

For Students & Learners

EXPLORE

Labs

Interactive Marimo notebooks. Change a parameter, see what breaks, build intuition.

Lab 17 · The Fleet Synthesis Explore

BUILD

TinyTorch

Build your own ML framework from scratch across 0 progressive modules. Zero magic.

tinytorch — tensor.py class Tensor: def __init__(self, data): self.data = data self.grad = 0.0 self._backward = lambda: None

MODEL

MLSys·im

First-principles performance modeling. One command, every bottleneck.

$ mlsysim eval Llama3_70B H100 --batch-size 1 mem-bound compute-bound b=1 b=32 b=128 Arithmetic Intensity FLOP/s

DEPLOY

Hardware Kits

Deploy ML to Arduino, Seeed, Grove, and Raspberry Pi. Real memory limits, real power budgets.

Arduino · Seeed · Grove · Raspberry Pi

For Career & Instructors

PRACTICE

StaffML

Physics-grounded interview questions for ML systems roles. Vault, drills, and mock interviews.

Systems Design L5 · Staff A 70B model needs 1,000 req/s. Walk through your hardware selection and parallelism strategy. Hardware Parallelism Trade-offs Cloud Edge Mobile TinyML

ADOPT

Instructor Hub

The AI Engineering Blueprint: two-semester syllabi, pedagogy guide, rubrics, and TA handbook.

The Blueprint — Course Architecture ML Systems · Two-Semester Curriculum Semester 1: Foundations 16 wks · Vol I · 8 assignments Semester 2: At Scale 16 wks · Vol II · capstone Assessment Rubrics · Peer review · Grading Teaching Staff Pedagogy · TA handbook READY

TEACH

Lecture Slides

0 Beamer decks with speaker notes and 0 original SVG diagrams. Drop in and teach.

Intro Systems DNN Training Accel Deploy Ethics The Iron Law of ML Systems T = D/BW + O/(R·η) + L Data Term — memory bandwidth Compute Term — utilization η ≤ 0.7 Latency Term — orchestration overhead Harvard University · ML Systems 12 / 38

FOLLOW

Newsletter

Updates on the curriculum, new chapters, and what the community is building.

MLSysBook Weekly 4 New: Vol II Ch. 7 — Fault Tolerance Updated: TinyTorch Module 0 Community: 1,200+ PRs merged Milestone: 28,000+ GitHub stars Join 1,100+ subscribers
Support the Mission
↓

OUR MISSION

AI education should be
free and open to everyone.

Everyone calls AI the new electricity — but electricity is useless without engineers who can build the grid. For AI to be efficient, reliable, and safe, the world needs engineers who understand how to build it.

That knowledge should be accessible to anyone willing to learn. This curriculum is our commitment to making it so.

Live readership — 18,389 readers a month on average across 204 countries and territories

Top countries: United States · India · Singapore · China · Vietnam

313,063
readers
204
countries &
territories
18,389
avg. a month

Our goal: 1,000,000 AI engineers by 2030

Star on GitHub 28,132

Every star helps another engineer find this book.
Our sponsors and partners keep it free for all of them.