Quantization resources worth your time
Primary sources first: the papers that introduced each method, the libraries that implement them, and courses that teach the maths properly.
Foundational papers
- arxiv.org
GPTQ: accurate post-training quantization for generative pre-trained transformers
Second-order, layer-wise weight quantization to 3–4 bits.
- arxiv.org
AWQ: activation-aware weight quantization
Protects salient channels found from activation statistics.
- arxiv.org
LLM.int8(): 8-bit matrix multiplication for transformers at scale
Mixed-precision decomposition that isolates outlier features.
- arxiv.org
QLoRA: efficient finetuning of quantized LLMs
4-bit NormalFloat, double quantization and paged optimizers.
- arxiv.org
SmoothQuant: accurate and efficient post-training quantization
Migrates activation outliers into weights to enable W8A8.
- arxiv.org
The era of 1-bit LLMs: all large language models are in 1.58 bits
Ternary-weight models trained from scratch (BitNet b1.58).
- arxiv.org
A survey of quantization methods for efficient neural network inference
Broad survey of uniform, non-uniform, PTQ and QAT methods.
- arxiv.org
Quantizing deep convolutional networks for efficient inference: a whitepaper
Classic practical guide to integer quantization and QAT.
- arxiv.org
Neural network quantization for efficient inference: a survey
Survey covering granularity, calibration and hardware.
Libraries and runtimes
- github.com
llama.cpp
C/C++ inference engine and home of the GGUF format and quantize tool.
- github.com
bitsandbytes
8-bit and 4-bit quantization used across the Hugging Face stack.
- huggingface.co
Hugging Face Transformers quantization docs
Overview and support matrix for every integrated method.
- docs.vllm.ai
vLLM quantization docs
Which quantized formats the vLLM server supports, by hardware.
- github.com
LLM Compressor
Library for producing GPTQ, AWQ, FP8 and other compressed checkpoints for vLLM.
- github.com
PyTorch torchao
PyTorch-native quantization and sparsity for training and inference.
- github.com
NVIDIA Model Optimizer
PTQ, QAT, FP8 and NVFP4 tooling for NVIDIA deployment stacks.
- onnxruntime.ai
ONNX Runtime quantization
Static and dynamic INT8 quantization for ONNX models.
- github.com
Intel Neural Compressor
Quantization and compression toolkit for Intel hardware and beyond.
- github.com
Apple MLX
Array framework for Apple silicon with built-in quantized model support.
- ollama.com
Ollama
Runs GGUF models locally with a simple CLI and API.
Courses and long reads
- efficientml.ai
MIT 6.5940 EfficientML.ai
University course on efficient deep learning, with full lecture videos.
- www.deeplearning.ai
DeepLearning.AI short courses
Free short courses, including quantization fundamentals and in-depth tracks.
- newsletter.maartengrootendorst.com
A visual guide to quantization
Diagram-heavy explainer of LLM quantization concepts.
- huggingface.co
Hugging Face Optimum: quantization concept guide
Affine quantization maths, calibration and granularity.
Signals, audio and physics
- www.analog.com
Analog Devices MT-001: the SNR = 6.02N + 1.76 dB formula
Derivation and limits of the ideal quantization-noise formula.
- www.dspguide.com
The Scientist and Engineer's Guide to DSP, chapter 3
ADC and DAC: quantization, dither and sampling.
- www.ti.com
TI Precision Labs: ADCs
Free video training series on converter specs and noise.
- phet.colorado.edu
PhET: models of the hydrogen atom
Interactive simulation of atomic models and spectra.
- physics.nist.gov
NIST CODATA fundamental constants
Authoritative values for h, e, m_e and the Rydberg constant.
Want this done for your model?
Tell us your model, hardware and target. We send back a free optimization plan: recommended format, expected memory and a quality-check method.