Quantization Fundamentals with Hugging Face

Quantization Fundamentals with Hugging Face is the best free, no-fluff first hour for beginners who want to actually run model-compression code, but it is too shallow for anyone past square one. This DeepLearning.AI short course, taught by Hugging Face engineers Younes Belkada and Marc Sun, teaches two practical ways to shrink LLMs and other models: linear quantization with the Quanto library and 'downcasting' to BFloat16 with the Transformers library. It is genuinely an introduction, not a deep dive: the hands-on notebooks cover data types (FP32/FP16/BF16), loading models in different precisions, and the basic linear-quantization formula (scale and zero-point), with code applied to real open-source models. On the Coursera-hosted version it holds a 4.8/5 rating, but from only 15 ratings, so the score is real but thin. For anyone who wants to understand what model quantization actually is and run working compression code in an afternoon, it is a high-value free starting point. Learners who already know quantization, or who want to build a quantizer from scratch and handle 4-bit/QLoRA-style methods, will find it too shallow and should go straight to the follow-on course.

Free, well-taught, and accurate for absolute beginners to model compression, but very short (~1 hour, 3 hands-on lessons) and deliberately surface-level. It is a clear 'take' for newcomers to quantization and a clear 'skip' for practitioners who already know the basics or need production-depth techniques (custom quantizers, 4-bit, QLoRA, GPTQ), who should jump to the deeper follow-up.

Best for: ML practitioners and engineers who already know Python and have some PyTorch experience, understand basic deep-learning concepts, and want a fast, hands-on first exposure to why and how models are quantized to reduce size and speed up inference. Ideal for people who will deploy or run open-source LLMs/vision models on limited hardware and want to start using Hugging Face Transformers + Quanto immediately.

Skip if: Complete programming beginners (it assumes Python and some PyTorch), people with no ML background, and, at the other end, engineers who already understand quantization or need production depth: building a quantizer from scratch, 4-bit / 2-bit methods, QLoRA, GPTQ, SmoothQuant, calibration, or quantization-aware training. Those users should take 'Quantization in Depth' instead.