MIT 6.5940: TinyML and Efficient AI Computing (Fall 2026)
by MIT HAN Lab
Song Han's MIT course on making models small and fast — pruning, quantization and LLM deployment, with five labs.
Overview
Taught by MIT associate professor Song Han — author of Deep Compression, MCUNet, AWQ and SmoothQuant, techniques that sit inside most production quantization stacks — 6.5940 is structured as four chapters across 25 lectures. Chapter I, Efficient Inference, covers deep-learning basics, pruning and sparsity (two lectures), quantization (two lectures), neural architecture search (two lectures), knowledge distillation, MCUNet for microcontrollers, and TinyEngine and parallel processing. Chapter II, Domain-Specific Optimization, moves to Transformers and LLMs, LLM quantization and deployment, LLM post-training, long-context LLMs, vision Transformers, GANs/video/point clouds, and diffusion models. Chapter III covers distributed training across two lectures plus on-device training and transfer learning; Chapter IV closes on advanced LLMs and agents before three sessions of final-project presentations. Five individual labs carry the hands-on work — pruning, quantization, neural architecture search, LLM compression, and LLM deployment on a laptop — and students must complete at least four of the five; a group final project is written up as a four-page NeurIPS-format report. MIT lists 6.191 Computation Structures and 6.390 Introduction to Machine Learning as hard prerequisites and de-registers students who lack them, which is a fair signal of the difficulty. The Fall 2026 edition begins 10 September 2026 and posts slides and recordings as it runs; complete lecture playlists from the Fall 2023 and Fall 2024 editions are already free on the MIT HAN Lab YouTube channel, so the material is usable today without enrolling.
At a Glance
- Topic
- ML
- Level
- Advanced
- Format
- Course
- Cost
- Free
- Duration
- 25 lectures over ~14 weeks plus 5 labs; ~6-8 h/week self-paced
- Provider
- MIT HAN Lab
- Hands-on
- Yes — code/exercises
- Certificate
- None
What You’ll Learn
- ✓Prune and sparsify neural networks, then measure the accuracy-latency tradeoff
- ✓Apply post-training quantization and quantization-aware training down to INT4 weights
- ✓Run neural architecture search to find hardware-efficient model variants
- ✓Quantize and deploy an LLM on a laptop using TinyEngine in the final lab
- ✓Reason about long-context LLM serving, KV-cache cost and attention efficiency
- ✓Parallelize training across GPUs with data, model and pipeline parallelism
- ✓Compress diffusion models and vision Transformers with domain-specific techniques
Highlights
- •Taught by the author of Deep Compression, MCUNet, AWQ and SmoothQuant
- •The Fall 2026 syllabus adds LLM post-training, long-context LLMs, and agents
- •Five graded labs end with running a compressed LLM on your own laptop
- •Prior editions' full lecture videos and slides are free on YouTube — no enrolment needed
- •Teaches the theory behind the quantization formats vLLM and llama.cpp actually ship
Who It’s For
Best For
- ✓ML engineers serving models under latency, memory or GPU-cost constraints
- ✓Engineers shipping models to edge devices, phones or microcontrollers
- ✓Inference engineers who already use AWQ or GPTQ and want to understand why they work
Prerequisites
- •Solid Python and PyTorch plus prior machine-learning coursework (MIT requires 6.390)
- •Computer-architecture fundamentals — memory hierarchy, throughput vs latency (MIT requires 6.191)
- •Comfort reading research papers; the lectures are paper-dense
FAQ
What is MIT 6.5940: TinyML and Efficient AI Computing (Fall 2026)?
MIT 6.5940 is Song Han's graduate course on efficient machine learning: how to shrink and accelerate models so they run on phones, microcontrollers and constrained GPUs instead of a datacentre. The Fall 2026 syllabus runs from pruning and quantization through LLM quantization and deployment, long-context serving, diffusion models and distributed training. Finish the five labs and you can compress an LLM and run it on your own laptop.
Is MIT 6.5940: TinyML and Efficient AI Computing (Fall 2026) free?
MIT 6.5940: TinyML and Efficient AI Computing (Fall 2026) is free to access.
What level is MIT 6.5940: TinyML and Efficient AI Computing (Fall 2026) for?
MIT 6.5940: TinyML and Efficient AI Computing (Fall 2026) is aimed at a advanced audience. Recommended background: Solid Python and PyTorch plus prior machine-learning coursework (MIT requires 6.390), Computer-architecture fundamentals — memory hierarchy, throughput vs latency (MIT requires 6.191), Comfort reading research papers; the lectures are paper-dense.
How long does MIT 6.5940: TinyML and Efficient AI Computing (Fall 2026) take?
Expect roughly 25 lectures over ~14 weeks plus 5 labs; ~6-8 h/week self-paced. Most learners work through it at their own pace.
What will I learn from MIT 6.5940: TinyML and Efficient AI Computing (Fall 2026)?
You'll learn: Prune and sparsify neural networks, then measure the accuracy-latency tradeoff; Apply post-training quantization and quantization-aware training down to INT4 weights; Run neural architecture search to find hardware-efficient model variants; Quantize and deploy an LLM on a laptop using TinyEngine in the final lab; Reason about long-context LLM serving, KV-cache cost and attention efficiency; Parallelize training across GPUs with data, model and pipeline parallelism; Compress diffusion models and vision Transformers with domain-specific techniques.
Topics
Sources
This page was written from 3 sources, 1 on domains other than hanlab.mit.edu.